JPH02277180A - Pattern learning device - Google Patents

Abstract

PURPOSE:To avoid a learning stop state by deciding a learning state based on the total correction value of a weight matrix and the evaluation value of error between the signal of an output layer and a teacher signal. CONSTITUTION:A feedforward neural circuit network comprises plural layers including an input layer storage part 1002 and an output layer storage part 1005 and performs the learning operations by an adverse transmission method. A weight matrix storage part 1011 is initialized by a random number. An input pattern is inputted via a terminal 1001 as the vector value and stored in the part 1002. Then all elements of the vector are sent to a matrix product computing part 1003, and the total correction value is obtained for the vector value stored in the part 1011. An error calculation part 1210 obtains the evaluation value from an error between the signal of the part 1003 and a teacher signal. The part 1003 decides the learning state based on the total sum of the evaluation value and correction value. When the learning state is discontinued, a matrix correction part 1013 changes the output signals received from each layer to prevent the learning stop state.

Description

[Detailed Description of the Invention] (Field of Industrial Application) The present invention relates to a pattern learning device that avoids a situation in which learning stops even though learning has not progressed sufficiently in a feedforward neural network. be.

(Prior Art) Conventionally, the correspondence between input cover turns and output signals has been described in the first document (“As5oative Memory” b
yT.

Kohonen: Springer-Verlag:
(1977), but the system shown in this document was a linear system and non-linear correspondence was not possible. On the other hand, in a feedforward neural network, a pattern is input, and an ideal output signal is input many times as a teacher.
ion t.

Computing with Neural Net
s” by R,P,Lippmann: IEEE,
ASSP: April 1987. pp, 4-2
By the back propagation learning method shown in 2), it is possible to learn the nonlinear correspondence between the human cover turn and the output signal.

(Problems to be solved by the invention) In the feedforward type neural network described above,
A weight matrix is calculated for the input signal vector, and the result is multiplied by a nonlinear function called a sigmoid function.
The process of obtaining the resulting vector value, further calculating a weight matrix, and 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 using random numbers, etc., and the weight matrix values are modified by back propagation learning to find an optimal value.

However, in conventional learning methods, a state occurs in which the output signal and the teacher signal do not match, and the weight matrix is not corrected even though learning has not finished (learning stopped state), and learning progresses. Sometimes it stops happening.

The purpose of the present invention is to monitor the above-mentioned learning stop situation,
- It is an object of the present invention to provide a pattern learning device that can avoid learning suspension and continue learning when learning suspension occurs.

(Means for Solving the Problem) A pattern learning device obtained by the first invention includes a feedforward neural network consisting of a plurality of layers including an input layer and an output layer that perform learning by back propagation method, A means for determining whether learning has stopped based on a result of calculating the sum of correction amounts of a weight matrix used in a neural network and an evaluation amount of an error between an output layer signal and a teacher signal; The present invention is characterized in that it includes means for modifying the output signal from each layer when it is stopped.

The pattern learning device obtained according to the second aspect of the present invention includes a feedforward neural network consisting of a plurality of layers including an input layer and an output layer that perform learning using the back propagation method, and a neural network calculated from the signals of each layer. means for determining whether learning has stopped based on the evaluation value of the amount of modification of the weight matrix used in the process and the evaluation amount of the error between the output layer signal and the teacher signal; and means for modifying the output signal of each layer.

The pattern learning device obtained according to the third aspect of the present invention includes a feedforward neural network consisting of a plurality of layers including an input layer and an output that performs learning using the back propagation method, and the amount of correction of a matrix used in the neural network. means for determining whether learning has stopped based on the result of calculating the sum of The present invention is characterized by comprising a modification portion and means for modifying a slope control parameter of a sigmoid function for obtaining an output signal in a layer other than the input layer.

The pattern learning device obtained according to the fourth aspect of the present invention includes a feedforward neural network consisting of a plurality of layers including an input layer and an output that performs learning using the back propagation method, and an internal neural network calculated from signals of each layer. A means for determining whether learning has stopped based on the evaluation value of the amount of correction of the weight matrix used in and the evaluation amount of the error between the output layer signal and the teacher signal, and , comprising means for modifying a signal in an input layer, and means for modifying a slope control parameter of a sigmoid function for determining an output signal in a layer other than the input layer.

The pattern learning device obtained according to the fifth aspect of the present invention includes a feedforward neural network consisting of a plurality of layers including an input layer and an output that performs learning using a back propagation method, and the amount of correction of a matrix used in the neural network. means for determining whether learning has stopped based on the result of calculating the sum of The present invention is characterized by comprising means for modifying, and means for modifying the range between the maximum value and the minimum value of a sigmoid function for obtaining an output signal in a layer other than the input layer.

The pattern learning device obtained by the sixth aspect of the present invention includes a feedforward neural network consisting of a plurality of layers including an input layer and an output that performs learning by backpropagation method, and a neural network calculated from signals of each layer. means for determining whether learning has stopped based on the evaluation value of the amount of correction of the weight matrix used in the process and the evaluation amount of the error between the output layer signal and the teacher signal; In this case, the method is characterized by comprising means for modifying a signal in an input layer, and means for modifying a range between a maximum value and a minimum value of a sigmoid function for obtaining an output signal in a layer other than the input layer.

(Operation) In the first invention, the learning stop situation is observed as a case where the sum of the correction amounts of the weight matrix becomes less than a certain value, and the difference between the output signal and the teacher signal is larger than a threshold value, and in that case, By modifying the output signal from each layer and eliminating the factors that cause the amount of modification of the weight matrix to be 0, it is possible to forcibly modify the weight matrix and avoid stopping learning.

In the second invention, instead of checking the learning stop state by the sum of the correction amounts of the weight matrix, 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, By observing the case where the estimated amount is less than a certain value and the difference between the output signal and the teacher signal is larger than the threshold value, learning stoppage can be detected without calculating the amount of correction of the weight matrix. In the case of stopping learning, by correcting the output signal from each layer and eliminating factors that cause the amount of correction of the weight matrix to be O, it is possible to forcibly correct the weight matrix and avoid stopping learning. .

In the third invention, the learning stop situation is observed as a case where the sum of the correction amounts of the weight matrix becomes less than a certain value, and the difference between the output signal and the teacher signal is larger than a threshold value, and in that case, By changing the parameter that controls the slope of the sigmoid function that calculates the output signal in the layer, the slope is made gentler, making it difficult for values of 0 and 1 to be output, and the upper and lower limits of the signal in the input layer are changed. By limiting the weight matrix to 0, it is possible to correct the weight matrix and avoid stopping learning by eliminating factors that cause the weight matrix correction amount to be 0.

In the fourth invention, instead of checking the total amount of modification of the weight matrix to determine the learning stop state, the modification 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. It is possible to detect learning stoppage without calculating the correction amount of the weight matrix by observing the case where the estimated amount is less than a certain value and the difference between the output signal and the teacher signal is larger than a threshold value. In case of stopping learning, the input layer −
By changing the parameter that controls the slope of the sigmoid function that calculates the output signal of the layer other than the By limiting by value, it is possible to forcibly modify the weight matrix and avoid stopping learning by eliminating factors that cause the amount of modification of the weight matrix to be 0.

In the fifth invention, the learning stop situation is observed as a case where the sum of the correction amounts of the weight matrix becomes less than a certain value, and the difference between the output signal and the teacher signal is larger than a threshold value, and in that case, when By changing the range parameter that controls the maximum and minimum values of the sigmoid function that calculates the output signal of the layer in , it is difficult to output values of 0 and 1, and the signal of the input layer is limited by the upper and lower limit values. , it is possible to eliminate the factors that cause the amount of correction of the weight matrix to become 0, correct the weight matrix, and avoid learning stoppage.

In the sixth invention, instead of checking the sum of the correction amounts of the weight matrix to determine the learning stop state, 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. Observe as a case where the estimated amount is below a certain value and the difference between the output signal and the teacher signal is greater than the threshold,
To detect stopping of learning without calculating the amount of correction of a weight matrix. When learning is stopped, the range parameter that controls the maximum and minimum values of the sigmoid function that calculates the output signal of layers other than the input layer is changed to make it difficult to output values of 0 and 1, and the signal of the input layer is changed. By limiting the value by an upper limit value and a lower limit value, it is possible to forcibly correct the weight matrix and avoid stopping learning by eliminating factors that cause the amount of correction of the weight matrix to be 0.

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

FIGS. 1 and 10 are block diagrams showing embodiments of the chisel turn learning device of the first invention. The learning device of the present invention can be applied to a feedforward pattern learning device with two or more layers, and FIG. 1 describes a feedforward pattern learning device with three layers.
The figure explains a two-layer pattern learning device, but
It is also applicable to pattern learning devices with layers or more.

First, with reference to FIG. 10, a description will be given of a two-layer embodiment of the pattern learning device according to the first aspect of the present invention. Learning is performed by repeatedly applying a human cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are provided will be described. A matrix storage unit 101 that stores a matrix used in the matrix product calculation unit in advance.
It is assumed that 1 is initialized by a random number.

First, feedforward processing for calculating an output value from a human power value will be explained. The human cover turn is manually input from the terminal 1001 as a vector value and stored in the input layer storage unit 1002. The input layer storage unit 1002 includes a register 2 that holds an NQ-dimensional vector, and all elements of the vector are sent to the matrix product calculation unit 1003. The manually input vector value is oio (1≦i≦NQ), and the matrix value held in the weight matrix storage unit 1011 is wij.
l (1≦i≦NQ, 1≦j≦N□), output layer storage unit 10
Assuming that the vector value held in 05 is 0''jl (1≦j≦N), the matrix product calculation unit 1003 calculates 0j1
=Σ(Oj1×Wij1)+θj1. (1) Perform the following calculation. However, at -1, oj is a bias value and is stored in the weight matrix storage unit 1011. The resulting OJ'' (1≦j≦N1) is calculated in the S function section 1004 as o'j' = (1+ tanh(oj''))/2
(2) Obtain o) 1 by calculation,
It is stored in the output layer storage unit 1005 as an output signal.

Next, the weight matrix correction process will be explained.

A teacher signal, which is an ideal value, for the output signal obtained by the above-described feedforward processing is sent to the terminal 102.
0 is input and held in the teacher signal storage unit 1009, and the value tj (1≦j≦Nl) and the value 0''jl in the output signal storage unit 1005 are transferred to the error signal calculation unit 1010.
Error signal d) as d'J'"EJO'J" (1≦to≦N1)
(3) and further output signal 0)
Evaluate the back propagation error from 1, dj1=d'j'Xoj1X(1oj)
By executing the formula (4), dj'' (1≦j≦
After calculating Nl), the value is converted to the matrix correction unit 101.
Transfer to 3. Furthermore, the error signal d31 is also transferred to the matrix correction amount calculation unit 1016 and used for observing learning stoppage.

In the matrix correction unit 1013, Wij1=Wij1+aXdj1×010(5)θ11
=θi''+bXdj1 (6) The weight matrix is corrected by the calculation, and the result is stored again in the weight matrix storage unit 1011. Here, a is a positive coefficient, and for example, each is set to 0.2 and 0. Although it can be set to 1, this value itself is not an essential problem.

The matrix correction amount calculation unit 1016 calculates the weight matrix correction amount (aX
dj1×0i0) is S=ΣE 1aXdj”XoHol
It is obtained by calculation that executes the formula (7). Here, I
The calculation Xl is a calculation to find the absolute value of X. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is determined by E=Σ(d'j
1)2(8), and furthermore, s<'rhi
(9)
E>Th2 If the condition expressed by the above equation (9) is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 1017 and the output layer correction unit 1018. is sent to modify the output signal of each layer. In the example, Thl
was set to 0.1, and Th2 was set to 0.3, but this is not an essential problem. If the above conditions are not met, it is assumed that learning is not in a stopped state, and the matrix correction amount calculation unit 1
No signal is output from 016 to the correction section of each layer.

Upon receiving the learning stop signal, the input layer correction unit 1017 reads out the value oi0 held in the input layer storage unit 1002, and calculates oi0=Min(Max(oj, L), U).
(10) The input value oi0 is corrected by performing the calculation and stored in the human storage unit 1002 again. This modification is as shown in FIG. 7. If the curve shown at 701 in FIG. 7 is set to the value before modification, the portion larger than U and the portion smaller than U of this curve are 702 and 703 in FIG. 7, respectively. By making the following corrections, it is possible to avoid input layer values from becoming values close to 0 or 1. Here, Min(x, Y) represents a function that selects the smaller of X and y, and Max(x, y
) represents a function that selects the larger of X and y.

Further, U represents the upper limit value of the human input signal and can be set to 0.9, for example. L represents the lower limit value of the human input signal and can be set to 0.1, for example. However, the values of U and L are 0
．． It is not limited to 9.0.1 and may be any value. Of course, U>L. The output layer correction unit 1018 also performs the above (1)
0) Execute the same calculation as in equation 0 to correct the output layer output value to avoid values close to 0 or 1, and to prevent learning from stopping. Then, from equation (3) again, equation (
By executing the calculation shown in 6, the weight matrix can be modified without stopping learning.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

Next, with reference to FIG. 1, a description will be given of an embodiment in which the pattern learning device of the first invention has three layers. Learning is performed by repeatedly applying a human cover turn and a teacher signal, but first a case will be described in which one input butter and one teacher signal are applied. A matrix storage unit 111.1 that stores a matrix used in the matrix product calculation unit in advance.
12 is initialized with a random number.

First, feedforward processing for calculating an output value from an input value will be explained. The input pattern is input as a vector value from the terminal 101 and stored in the input layer storage section 102. The input layer storage unit 102 is composed of registers (14) that hold No-dimensional vectors, and all elements of the vector are sent to the matrix product calculation unit 103. , wij the matrix value held in the first layer weight matrix storage unit 112
l (1≦i≦N011≦j≦Nl), intermediate storage unit 10
The vector value held at 5 is o'3'' (1≦j≦Nl)
Then, in the matrix product calculation unit 103, 0j1=Σ
The following calculation is performed: (Oj0×wij1)+θj1 (11). However, here oj1 is the bias value, and the first
The layer weight matrix is stored in the layer weight matrix storage section 112. Result o
jl (1≦j≦Nl) are each expressed as o'3''=(1+tanh(oj))/2 in the S function section 104.
(12) Obtain 0″j1 by calculation and store it in the intermediate layer storage unit 105.The weight matrix stored in the second layer weight matrix storage unit 11 is calculated as wjk2(l≦i
≦N1, l≦j≦N2), and if the bias value stored in the storage section 111 is 2 for θ, then the matrix calculation section 10
In step 6, calculate 0 to 2 = Σ vehicle 1 x Wij2) ten θ to 2 (13). The resulting 0k2 (1≦ and ≦N2) are each S
In the function section 107, o' is set to 2=(1+tanh(Ok2))/2
(14) Ok is determined by the calculation, and is stored in the output layer storage unit 108 as an output signal.

Next, back propagation processing for modifying the weight matrix will be explained.

A teacher signal, which is an ideal value, for the output signal obtained by the above-described feedforward processing is sent to the terminal 102.
The value t
k (1≦to≦N2) and the value ol in the output signal storage unit 108
k and are transferred to the error signal calculation unit 110, and the error signal d is
'k to d' 2 = tk - o' to 2 (1≦≦N2)
Calculate as (15) and further evaluate the back propagation error from the output signal 0' and 2, dk2=d' to 2Xo' to 2X (1-o+ to 2)
By executing the formula (16), dk2(
1≦ and ≦N2), the values are transferred to the matrix correction section 113 and the error backpropagation section 114. Furthermore, the error signal d°2 is also transferred to the matrix correction amount calculation unit 116 and used for observing when learning has stopped.

The matrix correction unit 113 corrects the second layer weight matrix by calculating 2 2 2 ′・1 (17) Wjk=Wjk+axdkxoJ θj2=θj2+bXdk2 (18) and stores the result in the second layer weight matrix storage unit 111 again. . Here, a and b are coefficients with positive values, and can be set to 0.2 and 0.1, respectively, as described above, but these values are not an essential problem.

dk2 calculated by executing equation (16) in the error calculation unit 110 is sent to the error backpropagation unit 114, and together with the weight matrix value transferred from the matrix storage unit 11, d31=Σ(Wjk2×dk2 )(19),
Furthermore, from 0′ and l transferred from the intermediate storage unit 105, 1. 1. 1 dj = d', XO,X (1-0'j")
By executing the formula (20), djl(1
≦j≦Nl), it is transferred to the matrix correction unit 115 and Wij1=Wij1+aXdj1×010(
21) Modify the first layer weight matrix by executing the equation θ11=θi”+b X djl (22), and store the result in the first layer weight matrix storage unit 112 again.

The matrix correction amount calculation unit 116 calculates the weight matrix correction amount (
The sum of axdk2×0)1 and axdj1×010) is calculated by executing the following formula. Here, the calculation lXI is a calculation to find the absolute value of X. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is expressed as E=Σ(d'k") using d', which is the result obtained by executing equation (15).
2 (24), and furthermore, S<Thl
(25)E
＞Th2 If the condition expressed by the above equation (25) is satisfied, it is determined that the learning is stopped, and
A signal is sent to the input layer correction section 117, the output layer correction section 118, and the output layer correction section 119 to correct the output signal of each layer.

If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 116 to the correction unit of each layer.

Upon receiving the learning stop signal, the input layer correction unit 117 reads out the value oi0 held in the input layer storage unit 102, and
0=Min(Max(oio, L), U)
(26) Execute the calculation to obtain the human power value oi
0 is corrected and stored in the human storage unit 102 again to avoid the value of the input layer from becoming a 0 or 1 value as explained in FIG. Here, Min(x, y) represents a function that selects the smaller of X and y, and Max(x, y
) represents a function that selects the larger of X and y.

Note that the values of U and L are as described above. The intermediate layer correction unit 118 and the output layer correction unit 119 also perform calculations similar to the above equation (26) to correct the intermediate layer output value to avoid values close to 0 or 1, and perform learning. Avoid causing outages. Then, again from equation (15) to equation (22)
By performing the calculation shown in , learning can be done without stopping,
Achieve weight matrix modification.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

Next, referring to FIG. 11, a description will be given of an embodiment in which the pattern learning device of the second invention has two layers.

Learning is performed by repeatedly applying human cover turns and teacher signals, but first a case will be described in which one large cover turn and one teacher signal ' are applied.

The matrix storage unit 1111 that stores matrices used in the matrix product calculation unit is initialized with random numbers in advance.

First, feedforward processing for calculating an output value from a human power value will be explained. The input cover pattern is input as a vector value from the terminal 1101 and stored in the input layer storage section 1102. The input layer storage unit 1102 is composed of registers that hold NO-dimensional vectors, and all elements of the vector are sent to the matrix product calculation unit 1103. The input vector value is oio (1≦i≦NQ), and the matrix value held in the weight matrix storage unit 1111 is wij.
l (1≦i≦No, 1≦j≦N), output layer storage section 11
The vector value held in 05 is O'J1 (1≦j≦Nl
), the matrix product calculation unit 1103 calculates 0j1
=Σ(oHoX w4j')+θj1
(27) Perform the calculation. Here, θj is a bias value and is stored in the weight matrix storage section.

The resulting ojl (1≦j≦Nl) are each
4, o'31=(1+tanh(ojl))/2
(28) Obtain o+jl by calculation and store it in the output layer storage unit 1105.

Next, the operation of the part that determines whether learning should be stopped will be explained.

A teacher signal, which is an ideal value, of the output signal obtained by the above-described feedforward processing is sent to the terminal 110.
2 and held in the teacher signal storage unit 1109, and transfers the value tj (1≦j≦Nl) and the values o+, l in the output signal storage unit 1105 to the error signal calculation unit 1110.
Error signal d) as d'31=t3-o'3' (1≦j≦Nl)
Calculate as (29). The result is sent to the matrix correction amount estimating unit 1116, which calculates the error evaluation amount E for determining whether to stop learning by executing the calculation expressed as E=Σ(d'j1)2(30).・ζ
do. Furthermore, the output layer storage unit 1105 or I? 0゜(1≦j≦
Nl) from the input layer storage unit 1102 to 010 (1≦i≦
NQ) is transferred to the matrix modification amount estimation unit 1116, and an estimated value S' of the matrix modification amount is calculated using the following formula.

s'=ΣΣ1(1-o'j1)Xo'j1Xo401
(31) J It is determined by a calculation that executes the formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal in the matrix correction amount estimating section 1116 is calculated as follows: 3'<Th1 E>Th2 (3
2) If the condition expressed by equation (32) above is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 1117 and output layer correction unit 1118. and modify the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 1116 to the correction unit of each layer.

This judgment condition is such that if the correction amount S of the weight matrix is close to 0 even though the error evaluation fiE is large, learning is considered to be stopped.

Upon receiving the learning stop signal, the input layer correction unit 1117 reads out the value oi0 held in the input layer storage unit 1102, and calculates oi0=Min(Max(oio, L), U).
(33) Execute the calculation to obtain the human power value o
i0 is corrected and stored in the input layer storage unit 1102 again,
Avoid input layer values close to 0 or 1. The output layer correction unit 1118 also executes calculations similar to the above equation (33), corrects the intermediate layer output and the output layer output value, and
or 1, which prevents learning from stopping.

Finally, the modified signal storage section 1102.110 of each layer
The back propagation process for modifying the weight matrix using the value of 5 will be explained.

The error calculation unit 1110 evaluates the back propagation error from d”jl obtained in advance according to equation (29) and the output signal o+jl held in the output signal storage unit, and calculates dj”=d’3”Xo’; 1X (1-Oj)
By executing the formula (34), dj
After calculating l (1≦j≦Nl), the matrix correction unit 1
112, Wij=Wij+axdj1×oi
0(35) θ11=θi” + b x djl(36
) is used to correct the weight matrix, and the result is stored in the weight matrix storage unit 1111 again.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

FIG. 2 is a diagram showing a three-layer feedforward type embodiment of the pattern learning device of the second invention.

Pattern learning is performed by repeatedly applying an input cover turn and a teacher signal. First, a case where one person cover turn and one teacher signal are applied will be described. In advance,
Matrix storage units 211 and 212 that store matrices used in the matrix product calculation unit are initialized with random numbers.

First, feedforward processing for calculating an output value from an input value will be explained. The human cover turn is manually entered as a vector value from the terminal 201 and stored in the input layer register/unique section 102. The input layer storage unit 202 is composed of a register that holds a No-dimensional vector, and all elements of the vector are sent to the matrix product calculation unit 203. Let the manually input vector value be 0iO (1≦i≦NQ), and let the matrix value held in the first layer weight matrix storage unit 212 be Wij
l (1≦i≦No, 1≦j≦Nl), intermediate layer storage unit 20
The vector value held in 5 is o'3" (1≦j≦N)
Then, in the matrix product calculation unit 203, oj1=Σ
The following calculation is performed: (O10×Wij1)+θj1 (37). However, here θj1 is the bias value, and the first
It is stored in the layer weight matrix storage section 212. Result 0
Each j1 (l≦j≦Nl) is expressed as o'3' = (1+ tanh (oj'))/2 in the S function section 204.
(38), 0) and 1 are obtained and stored in the intermediate storage unit 205. The weight matrix stored in the second layer weight matrix storage unit 211 is wjk2.
(1≦i≦N1.1≦j≦N2), and the storage unit 211
When the bias value stored in θ is 2, the matrix calculation unit 206 calculates 0k2=Σ(oj1×wij2)+θ=2(39). The result 0k2 (1≦≦N2) is converted to o' in the S function section 207 as 2= (1+ tanh(Ok2))/2.
(40), Ok is obtained by calculation and stored in the output layer storage unit 208 as an output signal.

Next, the operation of the part that determines whether learning should be stopped will be explained.

The output signal obtained by the above-mentioned feedforward processing is tagged, and the ideal value of the teacher signal is sent to the terminal 22.
0 and held in the teacher signal storage unit 209, and the value tk (1≦≦N2) and the value o in the output signal storage unit 208.
2 to + to the error signal calculation unit 210, and the error signal d'k to d' to 2=tk-o' to 2 (1≦≦N2)
Calculate as (41). The result is sent to the matrix correction amount estimating section, and the error evaluation amount E for determining whether to stop learning is calculated by executing the formula E = Σ (2 for d') 2 (42). . Furthermore, OK (1≦to≦N2) is sent from the output layer storage unit 208.
o'j' (1≦j≦N1) from the middle layer storage unit 205,
oio (1≦i≦NQ) is transferred from the input layer storage unit 202 to the matrix correction amount estimating unit 216, and an estimated value S' of the matrix correction amount is calculated using the following formula.

+ΣΣl(1o'j')Xo'31XoH01(43)
” is calculated by executing the formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal in the matrix correction amount estimating unit 216 is calculated as S'<Ti1l E>Th, 2 (
44) If the condition expressed by the above equation (44) is satisfied, it is determined that the learning is stopped, and the input layer correction unit 217, the intermediate layer correction unit 218, and the output A signal is sent to the pot main section 219 of layer 11 to correct the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 216 to the correction unit of each layer.

This determination condition is such that if the weight matrix correction amount S is close to 0 even though the error evaluation amount E is large, learning is considered to be stopped.

Upon receiving the learning stop signal, the input layer correction unit 217 reads out the value Oi held in the input layer storage unit 202 and sets it to oH0.
=Min(Max(oi, L), U)
(45) Perform the calculation to correct the input value Oi0 and store it in the input layer storage unit 202 again to avoid the input layer value from becoming a value close to 0 or 1. Middle layer correction section 21
8 and the output layer correction unit 219 also perform calculations similar to the above equation (45) to correct the intermediate layer output value to a value close to 0 or l, thereby eliminating the possibility of stopping learning.

Finally, a description will be given of back propagation processing in which the weight matrix is modified using the modified values of the signal storage units 202, 205, and 208 of each layer.

d)2 calculated in advance by the error calculation unit 210 according to equation (41) and the output signal 0! held in the output signal storage unit.
Evaluate the backpropagation error from 2 and dk2==d' to 2Xo' to 2X (1-o+ to 2)
(46) By performing the calculation, d
After determining k2 (1≦≦N2), the value is transferred to the matrix correction section 213 and the 8-point difference reverse transmission (crotch section 214).

Further, the error signal d' of 2 is also transferred to the matrix correction amount calculation unit 216 and used for observing when learning has stopped.

In the matrix correction unit 213, Wjk=Wjk+aXdkXOj(47)θj2=θj
The second layer weight matrix is corrected by calculation 2+b x dk2 (48), and the result is stored in the second layer weight matrix storage section 211 again.

The dk calculated by executing equation (46) in the error calculation unit 210 is sent to the error backpropagation unit 214, and together with the weight matrix value transferred from the matrix storage unit 211, d31=Σ(Wjk2×dk2 )(49),
Furthermore, from 0' and 1 transferred from the intermediate storage unit 205, "dj": d'j1XO'j"X (1-0'j1)
By executing the formula (50), djl
After finding (1≦j≦Nl), the matrix correction unit 21
5, Wij1=Wijl+axdjl×0i
0(51) θ11=θi” +b X djl(52)
The first layer weight matrix is corrected by performing calculation using the following formula, and the result is stored in the first layer weight matrix storage unit 212 again.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

Next, referring to FIG. 12, a description will be given of a two-layer embodiment of the pattern learning device according to the third aspect of the present invention. Learning is performed by repeatedly applying an input cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are applied will be explained. A matrix storage unit 121 that stores a matrix used in the matrix product calculation unit in advance.
1 is initialized with a random number.

First, feedforward processing for calculating an output value from an input value will be explained. The input cover pattern is input as a vector value from the terminal 1201 and stored in the input layer storage section 1202. The input layer storage unit 1202 is composed of registers that hold NQ-dimensional vectors, and all elements of the vectors are sent to the matrix product calculation unit 1203. Let the input vector value be oHo (1≦i≦NQ), and let the matrix value held in the weight matrix storage unit 1212 be W...
1 (1≦i≦NQ, 1≦j≦Nl), and the vector value held in the other unit 1205 that indicates the output layer is o'3'' (1≦
j≦Nl), the matrix product calculation unit 1203 calculates oj1=Σ(Oj0×wij1) 10j'
(53) Perform the calculation. However, here θ
j1 is a bias value and is stored in the weight matrix storage section. The resulting ojl (1≦j≦N) is calculated as o'31=(1+tanh(oj/UO))/2 in the S function section 1204, respectively.
(54) Obtain ojl by calculation,
It is stored in the output layer storage unit 1205 as an output signal. In the example, UO was initially set to 0.75.

Next, the weight matrix correction process will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 122.
The value tj (1≦j≦N1) and the value o+jl in the output signal storage unit 1205 are manually stored in the teacher signal storage unit 1209 from 0 and transferred to the error signal calculation unit 1210.
Error signal d'j is d'j1=t3-o'3' (1≦j≦Nl)
(55) and further output signal 0)
1 and evaluate the back propagation error, dj1=d'j'XO'j'X(1-0,>
By executing the formula (56), djl(1
≦j≦Nl), the value is transferred to the matrix correction unit 1213. Furthermore, the error signal dj1 is also transferred to the matrix correction amount calculation unit 1216 and used for observing learning stoppage.

In the matrix correction unit 1213, Wij1=Wij1+axdj1×0iO(57)θ1
1=-θi" + b Xdj"
The weight matrix is corrected by calculation (58), and the result is stored in the weight matrix storage unit 1211 again.

The matrix correction amount calculation unit 1216 calculates the weight matrix correction amount (a
S=ΣΣ1aXdj
1Xoi01 (59)'' l It is obtained by calculation that executes the formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is determined by executing the formula E=Σ(d'j')2(60) from d)1 of the result obtained by executing (123). Then, S<Thl
(61)
Execute the conditional judgment that E>Th2, and if the condition expressed by the above equation (61) is satisfied, it is determined that the learning is stopped, and
A signal is sent to the input layer correction unit 1217 and the output layer UO correction unit 1218 to recalculate the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 1216 to the correction unit of each layer.

Upon receiving the learning stop signal, the input layer correction unit 1217 reads out the value Oi0 held in the input layer storage unit 1202, and oH0=Min(Max(oio, L), U)
(62) The input value oi0 is corrected by performing the calculation and stored in the input storage unit 1202 again to prevent the input layer value from becoming a value close to 0 or 1 as shown in FIG.

In the UO correction unit 1218 in the output layer, UO=UOXc (6
3) Realize the calculation of the following formula, make the slope of the sigmoid function gentle as shown in Figure 8, recalculate the output layer output values, and avoid them from becoming values close to O or 1. Eliminate the possibility of learning suspension. In the example, the parameter C was set to 2 for convenience. Then, the weight matrix is corrected by executing the calculations shown in equations (55) to (58) again. After completing one learning session, the UO is reinitialized.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

FIG. 3 is a diagram showing an embodiment of a three-layer pattern learning device according to the third invention. Learning is performed by repeatedly applying an input cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are applied will be explained. Matrix storage units 311 and 312 that store matrices used in the matrix product calculation unit are initialized with random numbers in advance.

First, feedforward processing for calculating an output value from an input value will be explained. The human cover turn is input as a vector value from the terminal 301 and stored in the input layer storage section 302. The input layer storage unit 302 is composed of a register that holds a No-dimensional vector, and all elements of the vector are sent to the matrix product calculation unit 303. The input vector value is 010 (1≦i≦NQ), and the matrix value held in the first layer weight matrix storage unit 312 is w5j1 (
1≦i≦No, 1≦j≦Nl), and the vector value held in the intermediate layer storage unit 305 is o'3' (1≦j≦Nl), then the matrix product calculation unit 303 calculates 0j1=Σ( O
The following calculation is performed: i0×Wij1)+θj1 (64). However, here, θj1 is a bias value and is stored in the first layer weight matrix storage section. Each of the results 0J1 (1≦j≦N1) is calculated as o'3'' = (1+tanh (oj1/UO))/2 in the S function section 304.
(65) Find 0)1 by the calculation,
It is stored in the output layer storage unit 305. In the example, UO was initially set to 0.75.

If the weight matrix stored in the second layer weight matrix storage section 211 is wjk2 (1≦i≦N1.1≦j≦N2) and the bias value stored in the second layer weight matrix storage section 211 is θ, then the matrix calculation section 206 Then, 0k2=Σ(0)1XWij2)+θ2
(66) Perform the calculation. The result 0k2 (1≦≦N2) is converted to o' in the S function section 307 as 2= (1+ tanh(ok2)/UO)/2.
(67), 0' is calculated to be 2, and is stored in the output layer storage unit 308 as an output signal.

Next, back propagation processing for modifying the weight matrix will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 302.
The value t
k (1≦to≦N2) and the value ol in the output signal storage section 308
k2 and are transferred to the error signal calculation unit 130, and the error signal d'k is converted to d' by 2=tk-o' by 2 (1≦≦N2)
Calculate as (68) and further output signal OK
Evaluate the back propagation error from dk2 = 2X for d' 2X for g' (2 for 1-o')
After calculating dk2 (1≦≦N2) by executing the equation (69), the value is transferred to the matrix correction unit 313 and the error backpropagation unit 314. Furthermore, the error signal -5dk is calculated by the matrix correction amount calculation unit 3.
16 and used to observe learning stoppage.

The matrix correction unit 313 corrects the second layer weight matrix by calculating 2 · 1 (70) Wjk = Wjk + aXdkXO) Oj2 = -2 + bxdk2 (71), and stores the result in the second layer weight 7-trix storage unit 311 again. .

The dk calculated by executing equation (69) in the error calculation unit 310 is sent to the error backpropagation unit 314, and together with the weight matrix value transferred from the matrix storage unit 311, dk is calculated as 1, p, 1=Σ( Calculate wjk2×dk2) (72), and further calculate dj"=:d'j"Xo'j1X(1-Oj) from 0' and 1 transferred from the intermediate storage unit 305.
By executing the formula (73), djl
After determining (1≦j≦Nl), the matrix correction unit 31
5, Wij1=Wij1+aXdj1×oi
O(74) θi=θi 10bXdj'
The first layer weight matrix is corrected by calculating the equation (75), and the result is stored in the first layer weight matrix storage unit 312 again.

The matrix correction amount calculation unit 316 calculates the weight matrix correction amount (aXdk2×o′j1, aX dj”X oi
') is calculated by executing the following formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is calculated using the formula (68
) is obtained by executing d゛ by 2, E=Σ
It is obtained by executing the formula (2)2(77) for (d'), and further S<Thl
(78)
Execute the conditional judgment that E > Th2, and if the condition expressed by the above formula (78) is satisfied, it is determined that the learning is stopped, and
Input layer correction unit 317, output layer correction unit 318, output layer U
A signal is sent to the O correction unit 319 to recalculate the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 316 to the correction unit of each layer.

Upon receiving the signal to stop learning, the input layer correction unit 317 reads out the value Oi held in the input layer storage unit 302 and sets it to oi0.
=Min(Max(oi, L), U)
(79) The input value Oi is corrected by performing the following calculation and stored in the input storage unit 302 again to avoid the value of the input layer from becoming a value close to 0 or 1.

UO correction unit 318 in the intermediate layer and UO correction unit 3 in the output layer
In 19, UO=UOXc (80
), the curve 801.80 in Figure 8 is obtained.
As shown in 2, the slope of the sigmoid function is made gentler,
Recalculate the intermediate layer output and output layer output values, and make sure they are O or 1.
Avoid values close to , and prevent learning from stopping. In FIG. 8, the curve 801 has UO=0.75, the curve 802 has UO=1.5, and C=2. In the example, the parameter C was set to 2 for convenience. Then, the weight matrix is corrected by executing the calculations shown in equations (68) to (75) again.

At this point, it is possible to further determine whether to stop learning as shown in equation (76) to equation (78), and if equation (78) is still satisfied, it is also possible to correct UO again using equation (80). be. After completing one learning session, the UO is reinitialized.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

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

First, feedforward processing for calculating an output value from an input value will be explained. The input pattern is manually input as a vector value from the terminal 1301 and stored in the input layer storage section 1302. The input layer storage unit 1302 is composed of registers that hold NO-dimensional vectors, and all elements of the vector are sent to the matrix product calculation unit 1303. Let the input vector value be 010 (1≦i≦NQ), and let the matrix value held in the weight matrix storage unit 1311 be W...
1 (1≦i≦No, 1≦j≦N1), output layer storage section 13
The vector value held in 05 is O'J1 (1≦j≦N1
), the matrix product calculation unit 1303 calculates 0j1
=Σ(Oj0×Wij1)+θj1 (
81) Perform the following calculation. However, here, θj1 is a bias value and is stored in the weight matrix storage section.

Each of the results 0J1 (1≦j5N1) is the S function section 130
4, o'31=(1+tanh(oj'/UO))/2
(82), or and 1 are obtained and stored in the output layer storage unit 1305.

Next, the operation of the part that determines whether learning should be stopped will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 132.
0 and held in the teacher signal storage section 1309, and transfers the value (1≦j≦N□) and the values 0' and 1 in the output signal storage section 1305 to the error signal calculation section 1310. Error signal d) as d'3'=t3-o'4'' (1≦j≦Nl)
Calculate as (83). The result is sent to the matrix correction amount estimating unit 1316, which calculates the error evaluation amount E for determining whether to stop learning by performing the calculation expressed by the formula: E=Σ(d'j1)2(84). demand. Further, from the output layer storage unit 1305, 0”・1 (1≦j≦Nl
) from the input layer storage unit 1302, 0i0 (1≦i≦NQ
) is transferred to the matrix correction amount estimating unit 1316, and the estimated value S'' of the matrix correction amount is calculated using the following formula.

3'=ΣΣ1(1-o'31)Xo'31Xoi01
(85) 1'' is calculated by executing the formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal in the matrix correction amount estimation unit 1316 is calculated as follows: S'<Th1 E>Th2 (8
6) If the condition expressed by the above equation (86) 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. and modify the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 1316 to the correction unit of each layer.

This determination condition is such that if the weight matrix correction amount S is close to 0 even though the error evaluation amount E is large, learning is considered to be stopped.

Upon receiving the learning stop signal, the input layer correction unit 1317 reads out the value oiO held in the input layer storage unit 1302, and calculates oi0=Min(Max(oio, L), U).
(87) The input value oi0 is corrected by performing the following calculation and stored in the input layer storage unit 1302 again to prevent the input layer value from becoming a value close to 0 or 1. U at the output layer
The O correction unit 1318 calculates UO=UOX c (88
), and as explained using Figure 8, the slope of the sigmoid function is made gentler, and the intermediate layer output and output layer output values are recalculated, so that they are values close to 0 or 1. Avoid this and eliminate the possibility of learning suspension.

Then, the calculations shown in equations (81) and (82) are performed again, and the learning stop judgment shown in equations (83) to (86) is performed to determine whether equation (86) is still satisfied. If so, the UO is corrected again using equation (88).

Finally, the modified signal storage section 1302.130 of each layer
The back propagation process for modifying the weight matrix using the value of 5 will be explained.

The error calculation unit 1310 evaluates the back propagation error from d)1 calculated in advance according to equation (83) and the output signal 0″j1 held in the output signal storage unit, and calculates djl. :d”j1×0”j”X(1-0+,1)
By executing the formula (89), dj”
After finding (1≦j≦Nl), the matrix correction unit 13
Transfer to 12, wij"=wij'+aXdj"Xo
iO(90)θi1=θi”+bXdj”
The weight matrix is corrected by calculating the formula (91), and the result is stored in the weight matrix storage unit 1s again.
ii. After one learning session is completed, the UO is set to the initial value again.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

FIG. 4 is a diagram showing a fourth embodiment of the pattern learning device of the present invention in the case of three layers. Learning is performed by repeatedly applying an input cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are applied will be explained. Matrix storage units 411 and 412 that store matrices used in the matrix product calculation unit are initialized in advance with random numbers.

First, feedforward processing for calculating an output value from an input value will be explained. The input pattern is input as a vector value from the terminal 401 and stored in the input layer storage section 402. The input layer storage unit 402 is composed of a register that holds a No-dimensional vector, and all elements of the vector are sent to the matrix product calculation unit 403. If the input vector value is 010 (1≦i≦NQ) and the matrix is stored in the first layer weight matrix storage unit 412, the matrix product calculation unit 403 calculates Ojl:Σ(oj0×wij1) 〇j1 (92) Perform the calculation. However, here Oj
1 is a bias value, and the first layer weight matrix storage unit 4
12 is stored. The resultant ojl (1≦j≦Nl) is calculated as o'3'' = (1+ tanh (Oj''))/2 in the S function section 404, respectively.
(93) Obtain o) 1 by calculation,
The data is stored in the intermediate storage unit 405.

If the weight matrix stored in the second layer weight matrix storage section 411 is wjk2 (1≦i≦N1.1≦j≦N2) and the bias value stored in the second layer weight matrix storage section 411 is 2 for θ, then the matrix calculation section At 406, calculation is performed such that 0k2=E(oj1×Wik2)+θ becomes 2(94). The result 0k2 (l≦≦N2) is converted to o' in the S function section 407 by 2=(1+tanh(ok2))/2.
(95) is calculated to determine OK, and the result is stored in the output layer storage unit 408 as an output signal.

Next, the operation of the part that determines whether to stop learning will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 420.
The value t
k (1≦to≦N2) and the value 0′ in the output signal storage unit 408
2 and are transferred to the error signal calculation unit 410, and the error signal d'k is converted to d' to 2=tk-o' to 2 (l≦≦N2)
Calculate as (96). The results are sent to the matrix correction amount estimating section, and the error evaluation amount E for determining whether to stop learning is calculated by executing the formula E = Σ (2 for d') 2 (97). . Further, from the output layer storage unit 408, 2 (1≦≦N2) is set to 0′.
e, middle layer storage unit 405 empty O'J1 (1≦j≦N1
) from the input layer storage unit 402 o50 (1≦i≦NQ)
is transferred to the matrix modification amount estimating unit 416, and an estimated value S'' of the matrix modification amount is calculated using the following formula.

+ΣΣI(1-o'4')Xo'j"X oiol
(98) J It is determined by calculation that executes the formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal in the matrix correction amount estimation unit 416 is calculated as follows: S'<Th1 E>Th2 (9
9) If the condition expressed by the above equation (99) is satisfied, it is determined that the learning is stopped, and the input layer correction unit 417, intermediate layer correction unit 418, and output layer A signal is sent to a modification unit 419 to modify the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 416 to the correction unit of each layer.

This determination condition is such that if the weight matrix correction amount S is close to O even though the error evaluation amount E is large, learning is considered to be stopped.

Upon receiving the signal to stop learning, the input layer correction unit 417 reads out the value oi0 held in the input layer storage unit 402, and
=Min(Max(oio, L), U)
(100) is executed to correct the input value oi0 and store it in the input layer storage unit 402 again to prevent the input layer value from becoming a value close to 0 or 1.

UO correction unit 418 in the intermediate layer and UO correction unit 4 in the output layer
19, UO=UOX c (1
01), and as explained in Figure 8, the slope of the sigmoid function is made gentler, the intermediate layer output and the output layer output value are recalculated, and they are determined to be values close to 0 or 1. Avoid becoming a failure and causing unlearning. In the example, the parameter C was set to 2 for convenience. Then, the calculations shown in equations (93) and (95) are performed again, and the learning stop judgment shown in equations (97) to (99) is performed to determine whether equation (99) is still satisfied. If so, use the formula again (
101) to modify the UO.

Finally, a description will be given of back propagation processing in which the weight matrix is modified using the modified values of the signal storage units 402, 405, and 408 of each layer.

The output signal 0'' held in the output signal storage unit is d′, which is calculated in advance by the error calculation unit 410 according to equation (96).
Evaluate the back propagation error from k2, and calculate dk2 = d' to 2 x o' to 2 x α-o' to 2)
(102) By performing the calculation, dk
2 (1≦≦N2), the value is transferred to the matrix correction unit 413 and the error backpropagation unit 414. Furthermore, the error signal dk2 is also transferred to the matrix correction amount calculation unit 416 and used for observing learning stoppage.

The matrix correction unit 413 corrects the second layer weight matrix by calculating Wjk=Wjk+axdk×0) θj2=θj2+bxdk2 (104), and stores the result in the second layer weight matrix storage unit 411 again.

The dk calculated by executing equation (102) in the error calculation unit 410 is sent to the error backpropagation unit 414, and together with the weight matrix value transferred from the matrix storage unit 411, d)1=E(Wjk2Xdk2 ) (105), and further calculates dj1=d731xo'31x (1-o'3') from 0' and 1 transferred from the intermediate storage unit 405.
(106), dj
'(1≦j≦Nt), the matrix correction unit 4
15, wij1=Wij1+aXdj1×0
10(107)θ11=θi' +bX djl(10
8) Modify the first layer weight matrix by calculating the following formula, and store the result again in the first layer weight matrix storage unit 41.
Store in 2. When one learning session is completed, the UO is returned to its initial value.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

Next, referring to FIG. 14, a description will be given of a two-layer embodiment of the pattern learning device according to the fifth aspect of the present invention. Learning is performed by repeatedly applying an input cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are applied will be explained. A matrix storage unit 141 that stores a matrix used in the matrix product calculation unit in advance.
1 is initialized with a random number.

First, feedforward processing for calculating an output value from an input value will be explained. The input pattern is input as a vector value from the terminal 1401 and stored in the input layer storage section 1402. The input layer storage unit 1402 is composed of registers that hold NQ-dimensional vectors, and all elements of the vector are sent to the matrix product calculation unit 1403. Let the input vector value be oiO (1≦i≦NQ), and let the matrix value held in the weight matrix storage unit 1411 be wij
l (1≦i≦No, 1≦j≦Nl), output layer storage section 14
The vector value held in 05 is o'31 (1≦j≦N□
), in the matrix product calculation unit 1403, oj1
=Σ(oj0×Wij1)+θj' (
109) Perform the following calculation. However, here, Ojl is a bias value and is stored in the weight matrix storage section. The resulting OJ'' (1≦j≦N1) is each
In 04, o'31=p+(q-pX1+tanh(oj''))/
2 (110), o)1 is obtained and stored in the output layer storage unit 1405 as an output signal. In the embodiment, initially p.

q was set to O and 1, respectively.

Next, the weight matrix correction process will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 142.
The value tj (1≦j≦Nl) and the value o)1 in the output signal storage section 1405 are transferred to the error signal calculation section 1410 to calculate the error. The signal d'j is d) 1 = ぢ - o'j1 (1≦≦Nl)
(111) and further output signal 0'
Evaluate the back propagation error with j1, dj''=d'3'Xo'31X (1-o+, 1)
(112), dj'
After determining (1≦j≦Nl), the value is transferred to the matrix correction unit 1413. Furthermore, the error signal dj is also transferred to the matrix correction amount calculation unit 1416 and used for observing learning stoppage.

In the matrix correction unit 1413, Wij1=wij1+aXdj1×oi0(l13)θ
12θi +bXdj' (114
) is used to correct the weight matrix, and the result is stored in the weight matrix storage unit 1411 again.

The matrix correction amount calculation unit 1416 calculates the weight matrix correction amount (ax
dj1×oi0) is S=ΣΣla Xdj' Xoiol
(115)'' 1 is calculated by executing the formula. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is obtained by executing the formula E=Σ(d)1)2(116) from the result d3 obtained by executing (111), and further S<Thl
(117)
E>Th2 If the condition expressed by the above equation (117) is satisfied, it is determined that the learning is stopped, and the input layer correction unit 1417 and the output layer range correction unit 141
8 to recalculate the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 1416 to the correction unit of each layer.

Upon receiving the learning stop signal, the input layer correction unit 1417 reads out the value O1 held in the input layer storage unit 1402, and calculates o;0=Min(Max(oi, L), U).
(118) to correct the input value Oi and store it in the human power record storage unit 1402 again, so that the value of the input layer becomes a value close to 0 or 1 as shown in FIG. avoid.

In the range correction unit 1418 in the output layer, p=p'
(119) Realize the calculation of the formula q = q', correct the range of the sigmoid function as shown in Figure 9, and recalculate the output layer output value,
Avoid them from taking values close to 0 or 1, and prevent learning from stopping. In the example of FIG. 9, the curve 901
Then p=Q, q=1, and in curve 902 p:0.2, q=
It is 0.8. Then, the weight matrix is corrected by executing the calculations shown in equations (111) to (114) again. After completing one study, p
, q is reinitialized.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

FIG. 5 is a diagram showing an embodiment of the pattern learning device of the fifth invention in the case of three layers. Learning is performed by repeatedly applying an input cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are applied will be explained. Matrix storage units 511 and 512 that store matrices used in the matrix product calculation unit are initialized with random numbers in advance.

First, feedforward processing for calculating an output value from an input value will be explained. The human cover turn is input as a vector value from the terminal 501 and stored in the input layer storage section 502. The input layer storage unit 502 is composed of a register that holds a No-dimensional vector, and all elements of the vector are sent to the matrix product calculation unit 503. The input vector value is oio (1≦i≦NQ), and the matrix value held in the first layer weight matrix storage unit 512 is wHj'(
1≦i≦No, 1≦j≦N1), and the vector value held in the intermediate layer storage unit 505 is o'j1 (1≦j≦N1), the matrix product calculation unit 503 calculates oj1=Σ(O
The calculation is 10×Wij1)+0j1(120). However, here, θj1 is a bias value and is stored in the first layer weight matrix storage section. Result ojl(1
≦j≦Nl) is calculated as o'j'=p+(q-p)(1+tanh(oj'/U
O))/2 (121) or, 1
is obtained and stored in the output layer storage unit 505. In the embodiment, p and q were initially set to 0 and 1, respectively. If the weight matrix stored in the second layer weight matrix storage section 511 is wjk2 (1≦i≦N1.1≦j≦N2) and the bias value stored in the second layer weight matrix storage section 511 is θ, then the matrix calculation section 506 Now, calculate 0k2=Σ(Oj1×Wij2)+θ to 2(122). The resultant Ok2 (1≦≦N2) is converted to o' in the S function section 507 as 2=p + (p - q) (1 + tanh (O
k2)/UO)/2 (123) is used to determine OK, and the result is stored in the output layer storage unit 508 as an output signal.

Next, back propagation processing for modifying the weight matrix will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 520.
The value t
k (1≦to≦N2) and the value o+ in the output signal storage unit 508
2 and are transferred to the error signal calculation unit 530, and the error signal d'k is converted to d' to 2=tk-o' to 2 (1≦≦N2)
(124), and further output layer −”1
Evaluate the back propagation error from i-Ok, dk2 = 2X to d', 2X to o' (2 to 1-o')
After calculating dk2 (1≦≦N2) by executing the equation (125), the value is transferred to the matrix correction unit 513 and the error backpropagation unit 514. Furthermore, the error signal d' includes a matrix correction amount calculating section 51.
It is also transferred to No. 6 and used to observe when learning stops.

In the matrix correction unit 513, 2 2 2 +1 (126
)Wjk=Wjk+axdk×Oj θj2=θj2+ b
The layer weight matrix is stored in the layer weight matrix storage unit 511.

dk2 calculated by executing equation (125) in the error calculation unit 510 is sent to the error backpropagation unit 514, and together with the weight matrix value transferred from the matrix storage unit 511, d+j1=Σ4×2× dk2) (128),
Furthermore, it was transferred from the intermediate storage unit 505. ”・1 to dj' = d'j'xo'j1x (1-o'4')
By executing the formula (129), dj
After calculating l (1≦j≦Nl), the matrix correction unit 5
15, Wij1=Wij1+axdj1×0
iO(130)θ11=θi1+bXdj”
The first layer weight matrix is corrected by calculating the formula (131), and the result is again applied to the first layer weight matrix.
The layer weight matrix is stored in the layer weight matrix storage unit 512.

The matrix correction amount calculation unit 516 calculates the weight matrix correction amount (
a X dk2X o'j', aXdj1×010) is calculated by executing the following formula. Sara(ko,
The evaluation amount of the error between the output signal and the teacher signal is expressed by formula (124)
E=Σ(d
, ni 2, 2 (133)
It is obtained by executing the formula, and furthermore, S<Thl
(134)
E > Th2 is executed, and if the condition 1 expressed by the above equation (81) is satisfied, it is determined that the learning is stopped, and the input layer correction unit 517 and the intermediate layer UO correction unit 518, a signal is sent to the output layer correction unit 519 to recalculate the output signal of each layer. If the above conditions are not met) (it is assumed that there is no learning stop situation, the matrix correction amount calculation unit 516 does not output a signal to the correction unit of each layer) 1° Input layer correction unit 517 that received the learning stop signal Now, the input layer storage unit 502
Read the value oi0 held in 0! -Min(Ma
x (oi, L), U) (135)
This calculation is performed to correct the input value oiO and store it in the input storage unit 502 again to avoid the value of the input layer from becoming a value close to O or 1.

In the range correction unit 518 in the intermediate layer and the range correction unit 519 in the output layer, p=p' (13
6) Realize the calculation of the formula q"q', modify the sigmoid function range as explained in Figure 9, recalculate the intermediate layer output and output layer output value, and make sure that they are 0 or 1. Avoid values that are close to each other and prevent learning from stopping.After that, the weight matrix is corrected by executing the calculations shown in equations (120) to (131) again. After the learning is completed, p and q are initialized again.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

Next, with reference to FIG. 15, a description will be given of an embodiment in which the pattern learning device according to the sixth aspect of the present invention has two layers. Learning is performed by repeatedly applying a human cover turn and a teacher signal. First, a case in which one input cover turn and one teacher signal are provided will be described. The matrix storage unit 1511 that stores matrices used in the matrix product calculation unit is initialized with random numbers in advance.

First, feedforward processing for calculating an output value from an input value will be explained. The human cover turn is input as a vector value from the terminal 1501 and stored in the input layer storage section 1502. The input layer storage unit 1502 is composed of registers that hold NO-dimensional vectors, and all elements of the vector are sent to the matrix product calculation unit 1503. The manually input vector value is 01O (1≦i≦NQ), and the matrix value held in the weight matrix storage unit 1411 is VVI.
Jl (1≦i≦No, 1≦j≦N□), output layer record and other parts 1
The vector value held in 405 is O'J'' (1≦j≦N
l), the matrix product calculation unit 1503 calculates 0j
1=Σ(oj0×Wij1) 10j'
(137). However, here Oj1
is a bias value and is stored in the weight matrix storage section. The resulting OJ'' (1≦j≦Nl) is the S function part 1, respectively.
In 504, ol, 1=P + (q-P) (1+ tanh (Oj”
))/2 (138) or, l
is determined and stored in the output layer storage unit 1505 as an output signal. In the embodiment, initially p.

q was set to 0 and 1, respectively.

Next, the operation of the part that determines whether to stop learning will be explained.

For the output signal obtained by the above-described feedforward processing, the ideal value of the teacher signal is sent to the terminal 152.
0 is input and held in the teacher signal storage unit 1509, and the value chi (1≦j≦Nl) and the value o)1 in the output signal storage unit 1405 are transferred to the error signal calculation unit 1510 to calculate the error. Signal d) as dt, 1 = t, , 31 (1≦j≦N1)
Calculate as (139). The result is sent to the matrix correction amount estimating unit 1516, and the error evaluation amount E for determining whether to stop learning is determined by executing the calculation expressed by the formula E=Σ(d'j1)2(140). . Further, from the output layer storage unit 1505, 0”・1 (1≦j≦Nl
) from the input layer storage unit 1502, oio(1≦i≦NQ
) is transferred to the matrix modification amount estimation unit 1516, and the estimated value S' of the matrix modification amount is calculated using the following formula.

s'=MoΣl(1o'3')'Xo'3''Xoiol
(141)j is calculated by executing the equation. Furthermore, the evaluation amount of the error between the output signal and the teacher signal in the matrix correction amount estimating section 1516 is calculated as follows: S'<Th1 E>Th2 (14
2) If the condition expressed by the above equation (142) is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 1517 and the output layer correction unit 1418. and modify the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix correction amount calculation unit 1516 to the correction unit of each layer.

This determination condition is such that if the weight matrix correction amount S is close to 0 even though the error evaluation amount E is large, learning is considered to be stopped.

Upon receiving the learning stop signal, the input layer correction unit 1517 reads out the value O10 held in the input layer storage unit 1502, and calculates oi0=Min(Max(o70.L), U).
(143) Perform the calculation to correct the input value Oi0, and then store it in the input layer storage unit 150 again.
2 to avoid input layer values close to O or 1. The range correction unit 1318 in the output layer calculates p"p'q"q' (
144), and as mentioned above, make the slope of the sigmoid function gentler, recalculate the intermediate layer output and the output layer output value, and check that they become values close to 0 or 1. avoid causing learning failure. Then, the weight matrix is modified by executing the calculations shown in equations (156) and (158) again.After one learning session, p and q are initialized.Finally, the modified Signal storage section 1502.15 for each layer
The back propagation process for modifying the weight matrix using the value of 05 will be explained.

It is calculated in advance by the error calculation unit 1510 according to equation (156). Evaluate the back propagation error from p, 1 and the output layer -W OJ held in the output signal storage part, and dj'=d'31xo''j'X(1-o'j')
By executing the formula (145), dj”
After determining (1≦j≦Nl), the matrix correction unit 15
12, wij1=Wij1+aXdj1×0
iO(146)θ11=θi' + b X dj'
The weight matrix is corrected by calculating the equation (147), and the result is stored in the weight matrix storage unit 1511 again. After one learning session is completed, p and q are set to their initial values again.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

FIG. 6 is a diagram showing a sixth embodiment of the pattern learning device of the present invention. Learning is performed by repeatedly applying an input cover turn and a teacher signal, but first a case will be described in which one person cover turn and one teacher signal are applied. The matrix storage sections 61'l and 612 that store matrices used in the matrix product calculation section are initialized in advance with random numbers.

First, feedforward processing for calculating an output value from an input value will be explained. The input pattern is manually input as a vector value from the terminal 601 and stored in the input layer storage section 602. The input layer storage unit 602 is composed of registers that hold NQ-dimensional vectors, and all elements of the vector are sent to the matrix product calculation unit 603. Let the manually input vector value be oio (1≦i≦NQ), and let the Matonox value held in the first layer weight matrix storage unit 612 be wijl (
1≦i≦NQ, 1≦j≦Nl), and the vector value held in the intermediate layer storage unit 605 is o'3' (1≦j≦N4), then the matrix product calculation unit 603 calculates 0j1=Σ( O
The following calculation is performed: joXWijl)+ej1 (148). However, here, ejl is a bias value and is stored in the first layer weight matrix storage section 612. Result oj
Each l (1≦j≦Nl) is expressed as o'3'=p+(q-p)(l+tanh(Oj))/ in the S function section 604.
2 (148), the result is 0. seek,
It is stored in the middle layer storage unit 605. Let p and q be 1 and O.

If the weight matrix stored in the second layer weight matrix storage section 611 is wjk2 (1≦i≦N1.1≦j≦N2) and the bias value stored in the second layer weight matrix storage section 611 is θ, then the matrix calculation section 606 Then, ok2=Σ(Oj1×Wik2)+θ is 2'
(150). Result ok2(1
≦≦N2), respectively, 2=(1+t, anh(ok2))/2 for o' in the S function section 607.
Obtain olk by calculation (151),
It is stored in the output layer storage unit 608 as an output signal.

Next, the operation of the part that determines whether to stop learning will be explained.

A teacher signal, which is an ideal value, of the output signal obtained by the above-mentioned feedforward processing is sent to the terminal 620.
The value t
k (1≦to≦N2) and the value OK in the output signal storage unit 608
and is transferred to the error signal calculation unit 610 to obtain the error signal d'k
to d'2=tk-0''k2 (1≦≦N2)
(152). The result is sent to the matrix correction amount estimating section, and the error evaluation amount E for determining whether to stop learning is calculated by executing the formula E = Σ (2 for d') 2 (153). . Furthermore, 2 (1≦≦N2) from the output layer storage unit 608 to 0′
, from the middle layer storage unit 605 o'31 (1≦j system N1)
is transferred from the input layer storage unit 602 to 010 (1≦i≦NQ) to the matrix correction amount estimating unit 616, and the estimated value S′ of the matrix correction amount is calculated using the following formula.

+ΣΣ1(1-o'j1)xo'j'Xoi01
(154)j is calculated by executing the equation. Furthermore, the evaluation amount of the error between the output signal and the teacher signal in the matrix correction amount estimation unit 616 is calculated as follows: S'<Th1 E>Th2 (15
5) If the condition expressed by the above equation (155) is satisfied, it is determined that the learning is stopped, and the input layer correction unit 617, intermediate layer correction unit 618, and output layer A signal is sent to a modification unit 619 to modify the output signal of each layer. If the above conditions are not met, it is assumed that learning is not in a stopped state, and no signal is output from the matrix 1]3 positive quantity calculation unit 616 to the correction unit of each layer.

This determination condition is such that if the weight matrix correction amount S is close to 0 even though the error evaluation amount E is large, learning is considered to be stopped.

Upon receiving the learning stop signal, the input layer correction unit 617 reads out the value O10 held in the input layer storage unit 602, and
0= Min(Max(oHo, L), U)
(156) Execute the calculation and input value O
i0 is corrected and stored in the input layer storage unit 602 again to prevent the input layer value from becoming a value close to 0 or 1.

In the range correction unit 618 in the intermediate layer and the range correction unit 619 in the output layer, p=p'(1=(1' (
157), correct the sigmoid function range, recalculate the intermediate layer output and output layer output value,
Avoid them from taking values close to O or 1, and prevent learning from stopping. In the example, the parameters p',
q' was set to 0.2 and 0.8. Then, once again the formula (1
By executing the calculation shown in (49) and equation (151), the output value in each layer is corrected. After one learning session, p and q are initialized again.

Finally, a description will be given of back propagation processing in which the weight matrix is modified using the modified values of the signal storage units 602, 605, and 608 of each layer.

The output signal o stored in the output signal storage section is calculated according to equation (152) in advance by the error calculation section 610.
Evaluate the back propagation error from 2 to +, dk2 = 2X to d', 2X to o' (2 to 1-o')
After calculating dk2 (1≦≦N2) by executing the equation (158), the value is transferred to the matrix correction unit 613 and the error backpropagation unit 614. Furthermore, the error signal d' is also transferred to the matrix correction amount calculation unit 616 and used for observing whether learning has stopped.

In the matrix correction unit 613, Wjk=wjk+aXdkXOj(159)Oj2=θ
By calculating j2+ b X dk2 (160), the second
The layer weight matrix is modified and the result is stored again in the second layer weight matrix storage unit 611.

dk calculated by executing equation (158) in the error calculation unit 610 is sent to the error backpropagation unit 614, and together with the weight matrix value transferred from the matrix storage unit 611, d31=Σ(wjk2×dk2 )(161), and then calculate dj'=d'3'Xo'3'X(1-o'3'
(1
62), dj'(1≦j≦N
□) is transferred to the matrix correction unit 615, and Wij1=Wij1+axdj1×oi0(163)
θi1=θi' + b X dj
The first layer weight matrix is corrected by calculating the equation (164), and the result is stored in the first layer weight matrix storage unit 612 again. After completing one study session,
p and q are returned to their initial values.

The above is a description of one learning process, but by repeatedly performing this process, learning can be completed without being disturbed by stopping learning.

(Effects of the Invention) As is clear from the above explanation, in the present invention, learning stoppage is observed, and if learning stoppage occurs, the estimated amount of error between the output signal and the teacher signal becomes large. Nevertheless, if the amount of modification of the weight matrix is close to 0, learning stops. This is because O of the intermediate layer and output layer signals is approximately 1, and since the human input signal is close to O, the amount of correction of the weight matrix becomes 0. In this state, the learning device of the present invention prevents the intermediate layer and output layer signals from taking values close to 0 or 1, and prevents the input signals from taking values close to 0, thereby avoiding learning stoppage.

Using the learning device according to the present invention as the initial value of the weight matrix that has stopped learning by the conventional learning method,
As a result of learning the pattern, I was able to escape from the state where learning stopped and complete learning correctly.

[Brief explanation of drawings]

1 and 10 are block diagrams showing an embodiment of the first invention, FIGS. 2 and 11 are block diagrams showing an embodiment of the second invention, and FIGS. 3 and 12 are block diagrams showing an embodiment of the second invention. FIG. 4 and FIG. 13 are block diagrams showing an embodiment of the fourth invention, FIGS. 5 and 14 are block diagrams showing an embodiment of the fifth invention, FIGS. 6 and 15 are block diagrams showing an embodiment of the sixth invention. FIG. 7 is a diagram showing the effect of giving upper and lower limits to the output value of the sigmoid function to prevent the output value from approaching 1 or O. 8th
The figure shows the effect of preventing the output value of the sigmoid function from becoming close to 1 or O by changing the parameters for controlling the slope of the sigmoid function. FIG. 9 is a diagram showing the effect of changing the range of the output value of the sigmoid function to prevent the output value from becoming close to 1 or O. In the figure, 102, 202.302.402.502.602.1
002.1102.1202.1302゜1402, 1
502...input layer storage unit, 103, 203.303.
403.503.603.1003.1103.120
3.1303゜1403, 1503... Matrix product operation section, 104, 204.304.404.504.6
04.1004.1104.1204k. 1304, 1404.1504・S function part, 105, 2
05.305.405.505.605...Middle layer storage section, 106, 206.306.406.606...
Matrix product calculation unit, 107, 207.307.407
．． 607.607...S function part, 108, 208.3
08.408.508.608.1005.1105.
1205.1305°1405, 1505... Output layer storage section, 109, 209.309.409.509.6
09.1009.1109.1209.1309゜14
09, 1509...Teacher signal storage unit, 110, 210
．． 310.410.510.610.1010.111
0.1210.1310°1410, 1510...Error calculation section, 111, 211.311.411.511.
611.1011.1111.1211.1311゜1
411, 1511... Matrix storage section, 112, 2
12.312.412.512.612...Matrix Kokubu, 113, 213.313.413.513.6
13.1013.1113.1213.1313゜14
13, 1513... Matrix correction section, 114, 21
4.314.414.514.614...Error back propagation unit, 115, 215.315.415.515.615
...Matrix correction section, 116, 316.516.1
016.1116.1216.1316.1416.1
516... Matrix correction amount calculation unit, 216, 416.616... Matrix correction amount estimation unit, 117, 217.317.417.517.617.
1017.1117.1217.131? . 1417, 1517...input layer correction unit, 118, 21
8... Intermediate layer correction unit, 119, 219.1018.1118... Output layer correction unit, 318, 318.418.419.1218.13
18...UO correction department, 518, 519.618.61
9.1418.1518...Range correction section.

Claims (6)

[Claims] (1) A feedforward neural network consisting of multiple layers including an input layer and an output layer that performs learning using the back propagation method, and the results of calculating the total amount of modification of the weight matrix used within the neural network. , means for determining whether learning has stopped based on the evaluation amount of error between the output layer signal and the teacher signal, and means for correcting the output signal from each layer when learning has stopped. A pattern learning device characterized by: (2) A feedforward neural network consisting of multiple layers including an input layer and an output layer that performs learning using the back propagation method, and the amount of modification of the weight matrix used within the neural network calculated from the signals of each layer. means for determining whether learning has stopped based on the evaluation value and the evaluation amount of error between the output layer signal and the teacher signal; and means for correcting the output signal from each layer when learning has stopped. A pattern learning device comprising: (3) A feedforward neural network consisting of multiple layers including an input layer and an output that performs learning using the back-propagation method, the results of calculating the sum of the correction amounts of the matrix used in the neural network, and the output layer. means for determining whether learning has stopped based on the evaluation amount of error between the signal of the input layer and the teacher signal; 1. A pattern learning device comprising: means for modifying a slope control parameter of a sigmoid function for obtaining an output signal. (4) Evaluation of a feedforward neural network consisting of multiple layers including an input layer and an output that performs learning using the back propagation method, and the amount of modification of the weight matrix used within the neural network calculated from the signals of each layer. a means for determining whether learning has stopped based on the value and an evaluation amount of error between the output layer signal and the teacher signal; and a means for correcting the input layer signal when learning has stopped. , means for modifying a slope control parameter of a sigmoid function for obtaining an output signal in a layer other than the input layer. (5) A feedforward neural network consisting of multiple layers including an input layer and an output that performs learning using the back propagation method, the results of calculating the sum of the correction amounts of matrices used in the neural network, and the output layer. a means for determining whether learning has stopped based on the evaluation amount of error between the signal of the input layer and the teacher signal; a means for correcting the signal of the input layer when the learning has stopped; and a layer other than the input layer. 1. A pattern learning device comprising means for correcting a range between a maximum value and a minimum value of a sigmoid function for obtaining an output signal. (6) A feedforward neural network consisting of multiple layers including an input layer and an output that performs learning using the backpropagation method, and the amount of modification of the weight matrix used within the neural network calculated from the signals of each layer. means for determining whether learning has stopped based on the evaluation value and the evaluation amount of error between the output layer signal and the teacher signal; and means for correcting the input layer signal when learning has stopped. A pattern learning device comprising: and means for correcting a range between a maximum value and a minimum value of a sigmoid function for determining an output signal in a layer other than the input layer.