JPH0336659A - Learning controller for hierarchical neural network - Google Patents

Abstract

PURPOSE:To enable speedy and satisfactory convergence by extracting neuron, for which the size of an input is larger than a prescribed value, and executing the prescribed change of an output function. CONSTITUTION:In a learning device to repeat learning by back propagation and to converge a hierarchical neural network, neuron input f(X)=Xj(m) larger than the prescribed value extracted by a neuron extracting means is back propa gated to an m-layer control part Bm. This function f(X) is transformed to a function y=f(X/t)=[Xj(m)]/t, t>1 through a converter 11 to refer a derived function table 12. Accordingly, error to be back propagated is made large and correction quantity to coupled weight is also made large. Then, the network is speedily and satisfactorily converged to a desired state.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a learning control device for a hierarchical neural network, and more particularly, to a learning control device for a hierarchical neural network that can improve learning efficiency. .

[Prior Art] FIG. 13 shows a general network structure of a hierarchical neural network.

This hierarchical neural network 200 is a network in which a large number of neurons divided into a human layer, an intermediate layer, and an output layer are connected only in the direction from the human layer to the output layer. ~Vr.

The human resources layer is only the 1st layer, the middle layer is the 2nd layer to (L-4)
Up to this layer, the output layer is only the 2nd layer, and the number of neurons in the mth layer is N (+++).

Now, the neurons of the conventional human layer are the 5 neurons shown in Figure 14.
1, and its output function g is y 1 (1) = g (x 01)) = x 1 (1)
x 1 (1) = U 1 i = 1.2. . . . is an identity function expressed as N(1) (see Figure 20).

Here, 1(1) represents the i-th neuron in the first layer.

The conventional input layer neuron 51 may be constituted by a neuron device W181 as shown in FIG. 19, for example.

In this neuron device W81, the output function table 26 includes xi(
1) and y1(1) are stored as a table.

When the converter 25 is given the input xi(1), it refers to the output function table 26, obtains the output yi(1), and outputs it.

Conventional neurons in the intermediate layer and output layer have a configuration like 61 shown in Fig. 15, and the output function f is )' j(m)=f (x j(m))=(1+ax
p (-xj(o+)) l-'xj(+s)=ΣWi
j(a+-1)・)' i(+m-1)i=1.2.・
..., N(m-1) j=1.2. ..., N(m) Kaku = 2.3. . . . is a sigmoid function expressed by L (see Fig. 17).

Here, j(m) represents the jlth neuron in the mth layer. Moreover, Wij(m-1) represents the weight applied to the signal input from the first neuron of the m-1th layer to the j-th neuron of the m-th layer.

The conventional intermediate layer and output layer neurons 61 described above may be constituted by a neuron device 71 as shown in FIG. 16, for example.

In this neuron device 71, the multiplier 2 multiplies the weight Wij (+a-1) output from the weight register 53 by the output yi (+a-1) of the previous layer, and the multiplier 2 multiplies the signal input to the neuron device 71 by There are as many as.

Adder 4 adds the output signals of each multiplier 2, and obtains a total sum s
J(m) is input manually to the converter 5.

The converter 5 takes the sum sj (m) as the manual input xj (+s) to the table 6 of the output function, and y output from the table 6.
Output j(m).

The output function table 6 stores pairs of xj(m) and yj(+++) having a sigmoid function relationship shown in FIG. 17 as a table.

Learning control in this hierarchical neural network 200 is performed by the following pack propagation procedure.

(1) Give the first pattern of human power Ui to the human power layer.

(2) Output yj of each neuron from the input layer to the output layer
(m) is sequentially calculated to obtain the output yj(L) (=Vj).

(3) Based on the output yj (L) and the expected output dj, the error dj (L
).

dj(L)=(dj-yj(L))-t(xj
(L)) Here, t (xj(L)) is the value of the derivative of the output function at xj(L) (see FIG. 18).

(4) Modification amount ω1j(L
-1) is calculated.

ω1j(L-1)=c-dj(L)・y1(L-
1) However, C is a learning parameter.

(5) Modify the weight Wij(L-1) of the connection to the output layer.

Wij (L-1) (τ+1) = Wij (L-1) (τ) + ω1j (L-1) However, τ
is a discrete time, and (τ◆1) represents the next time after (τ).

(6) Error dj(
Find L-1).

dj (L-1) = Σ(δk(L)・Wjk(L-1
)(τ))・t (x j(L-1)) k=1.2. ..., N(L) The value Wjk(L-1)(τ) before modification is used as the connection weight.

(7) Similar to (4) and (5) above, the connection weight Wi
Modify j(L-2).

ω1j (L-2) = c ・dj (L-1) ・'j
1(L-2)W ij (L-2) (τ+1) =WiJ(L-2)(τ)+ω1j(L-2) (8) Below, repeat the same calculations as in (6) and (7) above. , the connection weight W i j (1) of each neuron in the second layer is modified.

(9) Repeat the above procedure by changing the pattern of input Ui and expected output dj. Do this for all necessary patterns.

(10) Calculate evaluation values E for all patterns.

E=Σ(d j−y j(L))”/2j=1.2
．． ..., N(L) (11) Repeat the above (1> to (10)) until the sum ΣE of the evaluation values E of all patterns becomes an appropriately small value E0 or less.

(12) Even if the number of learning times n (the number of repetitions of the above procedure) reaches the predetermined maximum value n0, if the sum ΣE of the evaluation values E does not become less than an appropriate small value E0, the connection weight Wij
(m) is randomly reset and the above procedure is performed again.

[Problems to be Solved by the Invention] In the learning control of the above-mentioned hierarchical neural network, the error δj (
m) is δj (m) = Σ(δk(a+1) ・W jk(m
) (r))・j (xj(m)) k=1.2. ..., N(L).

In this equation, the value of j (xj(m)) changes as shown in FIG. 18 depending on the value of xj(m) in the case of the output function f shown in FIG. 17, but the absolute value of xj(a+) If is large, the value becomes almost O, and the magnitude of the error δj (m) also becomes extremely small.

Therefore, the amount of modification of the connection weights ωij(m-1) also becomes extremely small, leading to the problem that the learning speed until convergence to the desired state becomes extremely slow, or that the learning speed does not converge to the desired state even if the number of times of learning is increased. There is a point.

Such a problem can be solved by the connection weight Wij(
It may be possible to solve this problem by randomly resetting m) and redoing pack propagation, but this is uncertain and inefficient.

Therefore, an object of the present invention is to provide a learning control device for a hierarchical neural network that can reliably and efficiently solve the above-mentioned problems.

[Means for Solving the Problems] A learning control device for a hierarchical neural network according to the present invention is a learning control device for converging a hierarchical neural network to a desired state by repeating learning by pack propagation. a neuron extraction means for extracting neurons whose input size is greater than or equal to a predetermined value, and an output function y of the extracted neuron;
= f (x) to y=f(x/t), t>1.

[Operation] In the hierarchical neural network learning control device of the present invention, for neurons with large inputs, the output function y
= f (x) becomes equivalently smaller.

Therefore, the value of the derivative also becomes a somewhat large value, and the error propagated by pack propagation also becomes a somewhat large value. Therefore, the amount of modification of the connection weights also increases,
This ensures faster convergence to the desired state.

[Example] Hereinafter, the present invention will be explained in more detail with reference to Examples shown in the drawings. Note that this invention is not limited to this.

FIG. 1 shows a hierarchical neural network 100 including a learning control device according to an embodiment of the present invention.

A learning pattern is input to neurons 1(t) to P(1) in the first layer from external inputs U1 to Up or from the first control unit B1. In addition, input X from the first control unit B1
and output y are input, and each output function table (
(see Figure 7).

Furthermore, neurons 1(1) to P(1) in the first layer are
The input x j (1) is output to the control unit B1. Outputs yl(1) to y of neurons 1(1) to P(1) in the first layer
pl(1) is manually input to neurons 1(2), . . . in the second layer in a predetermined combination, and is also input to the second control unit B2.

The neurons 1 (2), . . . in the second layer are connected to the second control unit B.
Since the weight W i j (1) obtained by dividing 2 by the temperature t is input, this is stored, and the weight W i j (1) obtained by dividing the stored temperature t is input in response to a request from the second control unit B2. and is output to the second control section B2.

The inputs of the second layer neurons 1 (2), . Further, the output Vj(2) is manually input to the third layer neurons 1(3), . . . in a predetermined combination, and is also manually input to the third control unit B3 (not shown).

The configuration from the third layer neuron to the second layer neuron is the same as the second layer neuron 1 (2), . . . . However, the outputs yj(L) of the neurons 1(L), .

The first control unit B1 corresponds to a supervisor that performs learning control, and controls neurons 1(1) to PI(1) of the first layer.
The output yj(
A deviation gj(L) between the expected output dj and the expected output dj is calculated and output to the second control unit BL.

gj(L)=d j−y j(L) Also, input layer! (1) Input the human power xj(1) of ~P(1), and set the output function of each neuron 1(1) to P(1) so as to normalize the output yj(1) with respect to these magnitudes. Change table settings.

The first control unit BL controls neurons 1 (L) to R of the second layer.
(L) deviation gj (L), one-man power xj (L), and the L-th
The output yi(
A new weight Wij(L-1)(・τ+1) is calculated from the current weight Wij(L-1)(τ) and the current weight Wij(L-1)(τ+1), and it is calculated based on the temperature t input from the first control unit B1. The division is performed to reset each neuron 1(L) to R(L).

Further, the deviation gj (L-1) of each neuron 1 (L-1), . . . in the L-1 layer is calculated and output to the L-1 controller BL-1.

gj(L-1)=Σ(δk(L)・Wjk(L-1)
(τ)) The configuration from the L-1st control section BL-1 to the second control section B2 is similar to the operation of the above-mentioned first control section BL.

The entire hierarchical neural network 100 has been described above, and each part will now be described in detail.

First, the neurons of the human power layer are constituted by a neuron device 21 shown in FIG.

In this neuron device M21, the adder 22 is provided in order to make it possible to accept a plurality of human inputs, but the hierarchical neural network 100 shown in FIG.
Since only one input signal corresponds to one neuron in the human power layer, only one human power of the adder 22 is used.

The output function table 24 includes x 1 (1) and y 1 (1
) are stored as a table. This xi(1)
The pair of and y1(1) is obtained from the minimum value Xai and maximum value Xbi of the human power x1(1) to the output function table 24, and the preferable minimum value Yai and maximum value Ybi of the output yi(1), ai=(Ybi-Yai)/(Xbi-Xai)b
i= (XbiYai −XaiYbi)/(Xb
i - Xai)y 1(1) = a ix 1(1) + b i is calculated by the first control unit B1 and set in the output function table 24. FIG. 8 is a conceptual diagram of the linear function expressed by the above equation.

The converter 23 uses the sum 5i(1) output from the adder 22 as an input x 1(1) to the output function table 24, converts it into an output yi(1), and outputs it.

Input Ui to the hierarchical neural network is, for example, an output signal from a light sensor, a weight sensor, an ultrasonic sensor, a video sensor, etc., and the values (levels, ranges) that these signals can take are various.

Therefore, as shown in FIG. 21, for example, there may be a case where the variation area of the human power Ul corresponds to a part of the variation area of the human power Ul, but the conventional first layer neuron device 81 shown in FIG. Then, the variation area of output yl(1) is output y
2(1), and there is a problem that the evaluation areas of input Ui and human power Ul are different. In addition, as shown in Fig. 22, if there is a bias in the variation range of input Ul and input U2, the variation range of output yl(1) and output y2(1) will also be biased, and the evaluation range will be biased. There are problems that arise.

For this reason, there are problems in that the learning speed required to obtain the correct output Vi in a hierarchical neural network may be slow, or the correct output Vi may not be obtained (no convergence) even if the number of times of learning is increased. However, in the neuron device 21, such problems have been solved.

That is, as shown in FIGS. 9 and 10, when the variation range of the human power Ul to the hierarchical neural network 100 corresponds to a part of the variation range of the human power U2, the input layer corresponding to the human power Ul The output function of the neuron Y 1 (
1) and the output function y2(1) of the neuron of the human-powered layer corresponding to the human-powered layer 2 are set as shown in the figure, and the output yl(
1) and y2(1) are made equal. Also, the first
As shown in FIGS. 1 and 12, the variation area of JUI entering the hierarchical neural network 100 and human power U2
If there is a bias in the fluctuation region of
(1) are respectively set as shown in the figure, and the variation areas of the output yl(1) and the output y2(1) are made equal.

This prevents the evaluation from being influenced by differences in the variation range of the manually input external signal.

Next, the neurons in the intermediate layer and the output layer are arranged as shown in FIG. 1.1.

In this neuron device 1, a multiplier 2 multiplies the weight Wij(m-1)/t output from the weight register 3 by the output yi(m-1) of the previous layer. There are as many signals as there are.

Adder 4 adds the output signals of each multiplier 2, and obtains a total sum s
j(m)/t is manually input to the converter 5.

The converter 5 is as described above! 9 j(m)/t is the human power xj(m) to table 6 of the output function, and yj(m) output from table 6 is output. Furthermore, xj(m) is output to the first control unit B1.

The output function table 6 stores pairs of xj(m) and yj(m) having the relationship of the sigmoid function f1 shown by the solid line in FIG. 3 as a table.

This output function fl is y j (11) = f 1 (x j (m)) = fl + e
xp(xj(m))l-' xj(m)=Σ(Wij(m-1)/t) ・
y i (m-1) i= 1.2. -, N(m-1) j=1.2. ..., N(m) m=2.3. ..., represented by L.

Next, the m-th opening control section Bm (m=L to 2) is configured as shown in the block diagram shown in FIG.

The converter 11 converts the human power x of the jth neuron in the mth layer
j(m) is input, and outputs the derivative f(xj(m)) of the output function by referring to the derivative table 12.

The multiplier 13 calculates the deviation εj (m) and the derivative f(xj
(m) and outputs the error δj (m).

This error δj (m) is stored in the δ memory 14 and output to the multiplier 15.18.

The multiplier 15 multiplies the output yi (m-1) of the m-1th layer by the error δj (m), and calculates the weight correction amount ωij (m
-1) is output.

The W memory 16 reads the weight Wij (m-1)/t held in the weight register 3 of each neuron, multiplies it by the temperature t by the multiplier 20b, and calculates the obtained weight Wij (m
-1) is remembered.

The adder 17 calculates the weight Wij(m-1)(r
) and the correction amount ωij(m-1) to output a new weight Wij(m-1)(τ+1).

The new weight WLj (m-1) (r 41) is calculated by the divider 2
0a is divided by the temperature t and output to the weight register 3 of the corresponding neuron.

The multiplier 18 calculates the error δj (m) and the weight W before correction.
ij(n+-1), and the accumulator 19 cumulatively adds the result. Therefore, the output of the accumulator 19 becomes the deviation εj(m-1) to the m-1th control unit Bm-1.

Next, the operation of the first control section B1 will be explained in detail with reference to FIGS. 6(a) and 6(b).

(0) The first control unit B1 receives a command from the operation unit 101, and at the same time, the target value EO of the evaluation value, the maximum value nO of the number of learning times, the initial value ΔtO1 for the increase in temperature t, the decrease for the temperature t The maximum value SO of the cumulative addition value Σxj(m) of the input xj(m) at each neuron is inputted at the rate.

(1) Number of learning times n=0. Temperature t=i and PI, P2)
, gives the first pattern input to the input layer (P3).

(2> Each neuron device 21 from the input layer to the output layer.
Sequentially calculate the output yj (gm) at 1, and calculate the output yj (L)
(=Vj) is obtained (P4).

At this time, x from each neuron device 1 of the intermediate layer and output layer
j(+s) are cumulatively added.

(3) The output layer (
The deviation sj (L) of the j-th neuron in the second layer) is calculated and output to the second control unit BL.

The second control unit BL calculates the error sj (L) by the following calculation (P5).

sj(L)= e j(L) −+ l (x j(L
)) (4) The first control unit BL to the second control unit B2 are operated in order, and the connection weight W i j (■) is corrected by pack propagation processing (P6).

(5) All necessary patterns (2) to (4) above
) is repeated (P7.P8).

(6) Apply the first pattern again (P9), calculate the output (Plo), and further calculate the evaluation value E (pH).

E=Σ(d j−y j(L))ri/2j=1,2.・
..., N(L) (7) Repeat (6) above for all necessary patterns (PI2.PI3).

(8) Increment the learning count n (P14).

(9) Calculate the sum ΣE of the evaluation values E for all patterns, and if ΣE5EO, end the learning and
If it is not ≦EO, move to PI6 (p. 15).

(10) Check whether the number of learning times n has reached the maximum value no, and if the number of learning times n has not reached the maximum value nO, proceed to P3. When the number of times of learning n reaches the maximum value no, it is determined that the state has converged to an undesirable state, and the process proceeds to relearning processing according to the present invention (P16).

The following operations are performed in the relearning process.

(11) Set the number of learning times n=o (Ql).

(12) Cumulative addition value Σxj(m) of each neuron device 1
and a preset maximum value SO, and if the former is smaller than the latter, the temperature t=1 is set, and if they are equal or larger, the temperature t=1+Δto- is set to -n. However, Δto<k
'n, the temperature is set to t=1 (Q2).

Initially, since n=o, the temperature t=1+Δto, which is greater than 1. As the number of learning n increases, the temperature t gradually increases.
will return to its original value of 1.

This temperature t is outputted to the second controller BL to the second controller 2.

The divider 20a divides the weight Wlj by this temperature t, which becomes a new weight, but this is essentially the same as s when t=1.
This is equivalent to dividing j(■) by t. Therefore, the output function fl(xj(m)) shown by the solid line in FIG. 3 has been substantially changed to the output function f 2 (s j (+a)) shown by the dotted line. In addition, the derivative i 1 (x j (m)) shown by the solid line in FIG. 4 is the derivative t 2 shown by the dotted line.
(s j (m)).

(13) Give the first pattern input to the input layer (Q3)
, each neuron device fil from the input layer to the output layer, 21
Sequentially calculate the output yj(a+) at and output yj(L)
(=Vj) is obtained (Q4).

(14) Determine the deviation sj(L) of the j-th neuron in the output layer (second layer) from the output yj(L) and the expected output dj, and output it to the second control unit BL.

The second control unit BL calculates the error sj (L) (Q5).

What should be noted here is that for neurons whose cumulative addition value Σxj(L) is small, x j(L) is
) = s j (1), and the error s j (L) is calculated as
(L) is calculated. Therefore, as can be understood from FIG. 4, in a neuron where the cumulative addition value Σxj(L) is large, even if 5j(L) is a large value?
2 (s j (m)) has a somewhat large value, and therefore the error s j (L) also has a somewhat large value.

The following Q6 to Q16 are the same processes as P6 to PI3 described above, so their explanation will be omitted, but in the pack propagation process (Q6), Sj (L) is Since a relatively large error Sj (L) can be obtained even with a large value, the connection weight correction amount ωij also becomes a relatively large value, increasing the probability of escaping from an undesirable convergence state (local optimum value).

Furthermore, as the number of times of learning n increases, the temperature t gradually returns to 1, making it easier to converge to a more desirable state (true optimal solution).

[Effects of the Invention] According to the learning control device for a hierarchical neural network of the present invention, it is possible to efficiently escape from an undesirable convergence state and converge to a more desirable state.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a hierarchical neural network including a learning control device according to an embodiment of the present invention, FIG. 2 is a block diagram of an example of a neuron device for the intermediate layer and output layer according to the present invention, and FIG. FIG. 2 is a conceptual diagram of the output function of the neuron device shown in FIG. 4, FIG. 4 is a conceptual diagram of the derivative of the output function of the neuron device shown in FIG. 2, FIG. 5 is a block diagram of the m-th control section, and FIG. (a) and (b) are flowcharts of the operation of a learning control device according to an embodiment of the present invention, FIG. 7 is a block diagram of an example of a neuron device in the human power layer, and FIG. 8 is an output of the neuron device shown in FIG. 7. Conceptual diagram of the function, Figures 9 to 12 are characteristic diagrams showing the relationship between the human power fluctuation range and the output fluctuation range, Figure I3 is a conceptual diagram of the hierarchical neural network, and Figure 14 is the neuron of the human power layer. Conceptual diagram, 1st
Figure 5 is a conceptual diagram of neurons in the intermediate layer and output layer.
Fig. 6 is a block diagram of an example of a conventional neuron device in the intermediate layer and output layer, Fig. 17 is a conceptual diagram of the output function of the neuron device shown in Fig. 16, and Fig. 18 is the output of the neuron device shown in Fig. 16. A conceptual diagram of the derivative of a function, FIG. 19 is a block diagram of an example of a conventional input layer neuron device, FIG. 20 is an input/output characteristic diagram of the neuron device shown in FIG. 19, and FIGS. 21 and 22 are FIG. 20 is a characteristic diagram showing the relationship between the input variation region and the output variation region in the neuron device shown in FIG. 19; (Explanation of symbols) 100...Hierarchical neural network Bl...
First control unit Bm...m-th control unit 1(1) to R(L)...neuron device...neuron device 2...multiplier 3...weight register 4...adder 5... ...Converter 6...Output function table 20a...Divider.

Claims (1)

[Claims] 1. In a learning control device that repeatedly performs backpropagation learning to converge a hierarchical neural network to a desired state, a neuron whose input magnitude is equal to or greater than a predetermined value is extracted from each neuron. Neuron extraction means;
Learning control for a hierarchical neural network, characterized in that it comprises an output function changing means for changing the output function y=f(x) of the extracted neuron to y=f(x/t), t>1. Device.