JPH04299443A - Neural network learning method - Google Patents

Abstract

PURPOSE:To attain an effective updating/learning in a neural network without storing the past teacher data. CONSTITUTION:When a learnt neural network is updated and learnt again, the data equivalent to the past teacher data are generated in the following method in place of a case where the past teacher data are stored. That is, a range of the input data available in the present neural network is sampled since the input/output characteristic of the neural network is secured from the past learning. Thus the input data are generated in an optional number of pieces, and the output data H are obtained in response to the data D. Then the sampled input/output characteristic is obtained. In such a constitution, the neural, network is learnt with use of the input data (D, A) including the input data A to be newly used for learning and the teacher date (H, B) including the teacher data B to be newly used for learning.

Description

Description: TECHNICAL FIELD The present invention relates to a learning method for neural networks. Specifically, the present invention relates to a method for further training an already trained neural network. [0002] A neural network is known as a circuit network that directly realizes causal relationships that are difficult to analyze theoretically through the learning effect of coupling coefficients. That is, the neural network learns the neural network coupling coefficients in advance for a plurality of discrete inputs so that the optimal output can be obtained for each input.
It is a circuit network that can directly obtain a valid output for any input. [0003] However, it is difficult to train a neural network so that it can respond to all possible events. Rather, when a new event occurs and an appropriate output cannot be obtained by actually using a trained neural network, learning that includes that event is usually performed. However, when a new neural network needs to be trained, it is not possible to perform the learning using only the new input data to be learned and the corresponding teacher data. This is because, as a result of the new learning, the previously learned items are deleted from the neural network. Therefore, when a new neural network is to be trained, it is necessary to train the neural network by including all the input data used in the previous learning and all the teacher data corresponding to the input data. [0005] As a result, it is necessary to save all input data and teacher data used in past learning, and as the neural network is used for a long time, the capacity of the storage device becomes enormous. Furthermore, since the amount of input data and teacher data used for learning becomes enormous, there is a problem that the learning time becomes long. The present invention has been made to solve the above problems, and its purpose is to perform efficient learning of a neural network that requires less storage capacity. [Means for Solving the Problems] The configuration of the invention for solving the above problems uses a trained neural network,
In a method for further learning using new input data and corresponding training data, multiple sets of arbitrary input data are input to a trained neural network, and multiple sets of output data corresponding to the input data are generated. , and multiple sets of input data including multiple sets of input data plus new input data to be learned, and multiple sets of output data plus new training data. A step is provided in which the neural network is trained using the data as training data. [Operation] When it becomes necessary to further train a neural network that has already been trained, a plurality of arbitrary discrete sets of input data representing as wide a range of events as possible are prepared. A trained neural network is activated for these multiple sets of input data, and corresponding multiple sets of output data are obtained. [0009]Next, a plurality of sets of input data are prepared by adding input data that needs to be further learned to the above-mentioned plurality of sets of input data. Furthermore, a plurality of sets of teacher data are prepared by adding teacher data corresponding to input data that needs to be trained to the output data of the trained neural network. Learning of the neural network is performed based on the plurality of sets of input data and the plurality of sets of teacher data. [0010] As described above, in the present invention, input data used in past learning and its teacher data are not stored or directly used. The input/output characteristics of a trained neural network are formed by past learning. By providing multiple sets of appropriate discrete input data to the neural network and obtaining multiple sets of output data at that time, it is possible to know the sampled input/output characteristics of the trained neural network. By learning by adding newly added input/output characteristics to the sampled input/output characteristics, effective update learning can be performed without erasing past learning items. [Example] 1. Neural Network As shown in FIG. 1, the neural network 10 of this embodiment has a three-layer structure including an input layer LI, an output layer LO, and a middle layer LM. The input layer LI has e input elements, the output layer LO has g output elements, and the middle layer LM has f
It has two output elements. A multilayer neural network is generally defined as a device that performs the following operations. The output Oij of the j-th element of the i-th layer is calculated by the following equation. However, i≧2. [Equation 1] Oij = f(Iij)

(1) [Math. 2] [Math. 3] f(x)=1/{1+exp(-x)}

(3) [0013] However, Vij is the j-th layer of the i-th layer.
The bias of the ith arithmetic element, Wi-1k,ij, is the bias of the ith arithmetic element.
The coupling coefficient between the k-th element of the -1 layer and the j-th element of the i-th layer, O1j represents the output value of the j-th element of the first layer. In other words, since it is the first layer, it outputs the input as it is without performing any calculations, so the input layer (
It is also the input value of the j-th element of the first layer). Next, a detailed calculation procedure of the three-layer neural network 10 shown in FIG. 1 will be explained with reference to FIG. 2. In step 100, the intermediate layer (second
For the j-th element of the layer), input the output value O1j (input data of the first layer) from each element of the input layer (first layer), and use equation (2) with the layer number and the first layer. A sum-of-products operation is performed using the following formula using the number of elements. [0015]Next, in step 102, the output of each element in the intermediate layer (second layer) is calculated by the sigmoid function of the product-sum function value of the input values of equation (4) using the following equation. Ru. The output value of the jth element of the second layer is calculated using the following equation. [Equation 5] O2j=f(I2j)=1/{1+exp(-I
2j) } (5) This output value O2j becomes the input value of each element of the output layer (third layer). Next, in step 104, a sum-of-products operation is performed on the input values of each element of the output layer (third layer). ##EQU00006## Next, in step 106, similarly to equation (5), the output value O3 of each element in the output layer is calculated using a sigmoid function.
j is calculated. [Equation 7] O3j=f(I3j)=1/{1+exp(-I3
j)} (7)
2. Creation of Already Trained Neural Network (Initial Learning) This neural network is trained in the procedure shown in FIG. 3 as initial learning. The coupling coefficients are performed by the well-known backpropagation method. This learning is repeatedly performed on a large number of input data related to various events so that each output becomes the respective optimal training data. At step 200 in FIG. 3, a learning signal for each element of the output layer is calculated using the following equation. [Formula 8] Y3j=(Tj-δj)・f'(I3j)
(8
) However, Tj is teacher data for an arbitrary output δj, and is given from the outside. Also, f'(x) is the derivative of the sigmoid function. Next, in step 202, learning data Y of the intermediate layer is calculated using the following equation. ##EQU9## Next, in step 204, each coupling coefficient of the output layer is corrected. The amount of correction is determined by the following formula. [Formula 10] Δω2i,3j(t)=P・Y3j・f(I2i)
+Q・Δω2i,3j(t-1) (10) However,
Δω2i,3j(t) is the t-th coupling coefficient between the j-th element of the output layer and the i-th element of the intermediate layer.
This is the amount of change in the second calculation. Also, Δω2i, 3j(t-
1) is the previous correction amount of the coupling coefficient. P,
Q is a proportionality constant. Therefore, the coupling coefficient is: [Formula 11] W2i,3j+Δω2i,3j(t) →W2i,
3j (11
), the corrected coupling coefficient is determined. Next, the process moves to step 206, where the coupling coefficient of each element in the intermediate layer is corrected. The amount of correction of the coupling coefficient is determined by the following equation, as in the case of the output layer. [Formula 12] Δω1i,2j(t)=P・Y2j・f(I1i)
+Q・Δω1i,2j(t-1) (12) Therefore, the coupling coefficient is: [Formula 13] W1i,2j + Δω1i,2j(t) →W
1i, 2j (
13), the corrected coupling coefficient is determined. Next, in step 208, it is determined whether one round of learning has been completed for all input data to be learned. If learning has not been completed for all input data, the process moves to step 210, and the next input data and the teacher data corresponding to the input data are set as learning target data. And step 200
The process returns to , and learning is performed on the next input data. In this way, once learning is completed for all input data, the determination result in step 208 becomes YES,
The process moves to step 212. In step 212, the coupling coefficient correction amount Δ
By determining whether ω has become less than or equal to a predetermined value, it is determined whether the coupling coefficient has converged. If the coupling coefficients have not converged, the process moves to step 214, where the first input data and the corresponding teacher data are set as the learning target data in order to perform the second learning on all input data. . Then, the process returns to step 200, and the above-described learning calculation is repeatedly executed. In this way, in step 212, the above learning calculation is repeatedly executed until the correction amount Δω of the coupling coefficient becomes less than or equal to a predetermined value and the coupling coefficient converges. As a result, a neural network initially trained on a wide range of initial events is completed. 3. Update Learning of Neural Network Next, the initially learned neural network is used for various types of actual input data. During this process, additional update learning is executed when a case arises where an appropriate output result cannot be obtained. The present invention is characterized by a method of performing such additional update learning. In step 300 of FIG. 4, input data D1, . . . , Dn, which has been randomly and discretely sampled widely in the past usage range, is prepared. This input data is defined as follows. Consider e pieces of data given to each of e pieces of input elements as one set of data. Then, an arbitrary p-th set of input data is represented by Dp, and input data for the j-th input element belonging to the set is represented by dpj. Dp represents a vector and dpj are the components of that vector. That is, D
p is defined by the following equation. [Formula 14] Dp = (dp1, dp2, ..., dpe-1,
dpe)
(14) Also, n sets of input data are D1, D2,
..., Dn-1, Dn. Hereinafter, all n input data groups will be referred to as input data group D. Next, in step 302, the input data D1,..., Dn are input to the neural network to obtain the corresponding output data H1,..., Hn, as shown in FIG. 5(a). . This output data is defined as follows. Regarding the output layer LO, consider the output data output from each of the g output elements as one set of data. Then, an arbitrary p-th set of output data is represented by Hp, and output data for the j-th output element belonging to the set is represented by hpj. Hp represents a vector, and hpj are the components of that vector. That is, Hp is defined by the following equation. [Formula 15] Hp = (hp1, hp2, ..., hpg-1,
hpg)
(15) Also, the output data of n sets are H1, H2,
..., Hn-1, Hn. Hereinafter, all n output data groups will be referred to as output data group H. Next, in step 304, (
As shown in a), m sets of input data related to events that require new teaching and m sets of teacher data corresponding to the respective input data are prepared. This input data is also represented by Ap using a vector, and its components are represented by apj. That is, Ap is defined by the following equation. [Formula 16] Ap = (ap1, ap2, ..., ape-1,
ape)
(16) Furthermore, if the number of sets of input data is expressed as m, then m sets of input data regarding a new event that requires teaching are expressed as A1, A2, . . . , Am. Hereinafter, the m input data groups will be referred to as input data group A. Further, the set of teacher data corresponding to the arbitrary input data Ap is represented by Bp using a vector, and its components are represented by bpj. That is, Bp
is defined by the following equation. [Formula 17] Bp = (bp1, bp2, ..., bpg-1,
bpg)
(17) Also, m sets of input data are A1, A2,
..., m sets of training data corresponding to Am are B1, B2,
..., expressed as Bm. Hereinafter, the m sets of teacher data will be referred to as a teacher data group B. Next, in step 306, (
As shown in b), (n+m) data groups (D1, ..., Dn,
A1,..., Am) are input data, and (n+m) data groups (H1,..., Hn,
B1,...,Bm) are used as teaching data. Then, similar to the initial learning described above, learning is performed on the input data and teacher data according to the flowchart of FIG. 3. First learning cycle If learning is performed in the order from input data group D to input data group A, in the first learning cycle, the output data of the neural network for input data (D1,...,Dn) is (H1,...,Hn). Therefore, in equation (8), the difference between the output data and the teacher data (Tj - δj in component representation) is zero, so all learning signals are zero. Therefore, regarding the learning of data group D, no change in the coupling coefficient is observed. Next, learning of input data group A is performed. At this time, (A1,..., Am)
The output data of the neural network with input data is not equal to the teacher data (H1,...,Hn). That is, in equation (8), for any input data Ap, the difference between the output data and the teaching data (in component representation, T
j−δj) is not zero. Therefore, learning data is generated based on this difference, learning is performed, and the coupling coefficient is changed. Second learning cycle In the second learning cycle, the coupling coefficients of the neural network have been changed by learning the input data group A, so the learning of the input data group (D1,...,Dn) , the output data of the neural network for any input data Dp is not equal to the teacher data Hp. That is, in equation (8), since the difference between the output data and the teacher data (Tj - δj in component representation) is not zero, learning data Y is generated and the coupling coefficient is changed. In this way, in the second and subsequent learning cycles, the input data group (D1, ..., D
Learning for n ) will also be performed. [0036] Multiple learning cycles The above learning cycle will be repeated multiple times.
In the end, the neural network uses the combined input data group (D1,...,Dn,A1,...,Am) and the combined teacher data group (H1,...,Hn,B1,...,Bm).
Learning will be performed regarding. Therefore, events learned in the past (events enveloped by input data group D1,...,Dn) and newly learned events (input data group A1
, ..., Am) can be formed. Furthermore, when update learning of the neural network becomes necessary, data group D is obtained through sampling using the latest neural network as described above.
What is necessary is to create the update learning shown in FIG. 4, which takes into account the newly added data group A. In this way,
It becomes possible to perform update learning one after another without storing the pairs of input data and teacher data used in past learning. Note that the number m of input data to be added may be one.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram showing the configuration of a neural network according to a specific embodiment of the present invention.

FIG. 2 is a flowchart showing the calculation procedure of the neural network according to the same embodiment.

FIG. 3 is a flowchart showing a learning procedure of the neural network according to the same embodiment.

FIG. 4 is a flowchart showing a procedure for update learning of the neural network according to the same embodiment.

FIG. 5 is a block diagram showing the concept of update learning of the neural network according to the same embodiment.

[Explanation of symbols]

10...Neural network LI...Input layer LM
...Middle layer Lo ...Output layer

Claims (1)

[Claims] [Claim 1] A trained neural network,
In a method for further learning using new input data and corresponding training data, inputting a plurality of sets of arbitrary input data to the trained neural network and outputting a plurality of sets corresponding to the input data. a step of obtaining data; and input data is a plurality of sets of data obtained by adding the new input data to be newly learned to the plurality of sets of input data, and the training data to be newly learned to the plurality of sets of output data. A method for learning a neural network, comprising the step of causing the neural network to learn using a plurality of sets of data including the above as training data.