JPH117435A - Neural network - Google Patents

Abstract

PROBLEM TO BE SOLVED: To efficiently obtain a high evaluation value by arranging two kinds of neural networks(NNs) in parallel at the time of executing learning and recognition based on the NNs, mutually comparing both the NNs in each divided learning and efficiently advancing the leaning of the NN having a better condition. SOLUTION: An NN group 8 consisting of an advancing side NN and a comparing side NN is formed. Final learning frequency and divided learning frequency are set up in a condition setting table 7. An NN control part 1 repeates processing for referring to the table 7, applying NN learning processing to respective NNs, comparing the advancing side NN with the comparing side NN in each divided learning frequency, selecting the NN having a high evaluation value, and advancing the learning. The initial value or condition of the NN can be set up by a random number or the like in each processing.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network type pattern discriminating apparatus and value estimating apparatus for information processing and learning means for the same.

[0002]

2. Description of the Related Art As a structure and learning means of a neural network (multi-layer hierarchical neural network), there is a typical error back propagation method (back propagation method).
(For example, “Neural Network Model and Connectionism”, written by Shunichi Amari, published by The University of Tokyo Press.) This outline will be described with reference to the unit internal state diagram of FIG. The relationship is (1)
Expressions (2) and (3) show that the input to the unit i is represented by Oj (j = 1 to N), and the coupling coefficient (weight) for each Oj is represented by Wij. At this time, the coupling coefficient Wij is initialized with random numbers within the specified range.

The sum of products of inputs Xi = ΣWijOj (1) The equation (1) is applied to a function f (Xi) for conversion. In general, a sigmoid function such as the following equation (2) that is differentiable is generally used as a function. f (Xi) = 1 / {1 + exp (−Xi)} (2) Output Yi = f (Xi) (3) where Yi The value is a number between 0 and 1. On the other hand, the units in the input layer use the input value as it is as the output value.

The error backpropagation method is a learning method for minimizing the square error of a unit between an output pattern Yi obtained from a learning pattern Xk and a desired output value Ii (hereinafter referred to as a teacher pattern) for the learning pattern. . Here, Figure 2
An example of such a three-layer structure will be described, but the same applies to a multilayer structure. The learning of the coupling coefficient is performed as follows. First, in the learning of the output layer, if the loss function is E (square error),
Where i is the unit number of the output layer, j is the unit number of the intermediate layer, k is the unit number of the input layer, and Oj
Is an output of the intermediate layer, Wij is a coupling coefficient between the intermediate layer and the output layer, Yi is an output of the output layer, and Ii is a teacher pattern. E (W) = 1 / 2Σ (Yi−Ii) · (Yi−Ii) (4) W indicates all Wij.

When the steepest descent method (probabilistic descent method) is applied, the coupling coefficient may be changed as shown in equation (6) using equation (5) as a learning signal. δE / δWij = (Yi−Ii) f ′ (ΣWijOj) Oj (5) New Wij = old Wij−cδE / δWij (6) where c is a learning coefficient, Oj is the output of the hidden layer.

The sigmoid function of the equation (2) can be generally expressed by the following equation (7), and by differentiating the sigmoid function, the original function is used and expressed by the equation (8). f (x) = 1 / {1 + exp (−x)} (7) f ′ (x) = f (x) (1−f (x)) (8) Since f (@WijOj) = Yi in practice, (8)
The equation is represented by f ′ (ΣWijOj) = Yi (1−Yi), and equation (5) becomes equation (9). δE / δWij = (Yi−Ii) Yi (1−Yi) Oj (9) Therefore, if c = 1, the expression (6) is expressed as follows: New Wij = Old Wij− (Yi−Ii) Yi ( 1-Yi) Oj (10) Here, D1i = (Ii−Yi) Yi (1
-Yi), equation (10) is represented by equation (11). New Wij = Old Wij + D1Oij (11) If learning of the hidden layer is E (square error) as the loss function,
It is expressed by equation (12). E (V) = １ ／ (Yi−Ii) · (Yi−Ii) (12) V is all Vjk. Here, Vjk represents a coupling coefficient between the intermediate layer and the input layer. When the steepest descent method (probabilistic descent method) is applied, the coupling coefficient may be changed as shown in Expression (14) using Expression (13) as a learning signal.

ΔE / δVjk = δE / δOj · δOj / δVjk = Σ (Yi-Ii) δYi / δOj · δOj / δVjk = ｛Σ (Yi-Ii) f ′ (WijOj) Wij｝ × f ′ (ΣVjkXk) Xk (13) From new Vi = old Vi−cδE / δVjk (14) Expression (14) is converted using Expressions (15) to (18),
Equation (19) is obtained. Dj =-{(Yi-Ii) f '(WijOj) Wij} = {D1iWij ... (15) f' (@ VjkXk) Xk = Oj (1-Oj) Xk ... (16) δE / δVjk = −DjOj (1−Oj) Xk (17) where Xk is an output of the input layer. Therefore, if c = 1, new Vjk = old Vjk + DjOj (1-Oj) Xk (18) where D2j = DjOj (1-Oj). Therefore, new Vjk = old Vjk + D2jXk (19) In this way, the learning equation of the coupling coefficient between the input layer and the intermediate layer can also be expressed. The calculation of the coupling coefficient starts from the unit of the output layer, moves to the unit of the intermediate layer, and then to the intermediate layer of the preceding stage. Therefore, the learning proceeds as follows. First, a learning pattern is input, and the result is calculated. All coupling coefficients are changed so as to reduce the error from the resulting teacher pattern. The learning pattern is input again. This is repeated until convergence. As the learning method, there are a sequential learning method for updating the coupling coefficient described above for each learning pattern and a collective learning method for correcting errors of all learning patterns collectively.

[0008]

The prior art has the following disadvantages. (1) In the conventional multilayer hierarchical structure type neural network, the initial values of the coupling coefficient between the input layer and the intermediate layer and the coupling coefficient between the intermediate layer and the output layer are set by random numbers, and the coupling coefficient is learned by learning data. Is adjusted, the learning does not proceed depending on the setting of the initial value of the coupling coefficient, and the learning does not converge even if the number of times of learning is increased. (2) Since the number of units in the intermediate layer in the conventional multilayer hierarchical structure type neural network does not have a fixed theoretical optimum number of units, it is set based on experience and intuition, and each coupling coefficient is adjusted by learning. Depending on the setting of the intermediate layer, learning does not proceed, and learning does not converge even if the number of times of learning is increased. (3) Since various learning parameters in the conventional multilayer hierarchical structure type neural network have different conditions for each task, they are set based on experience and intuition, and learning is performed to adjust each coupling coefficient. Depending on the setting, learning does not proceed, and learning does not converge even if the number of times of learning is increased.

SUMMARY OF THE INVENTION The present invention has been made in consideration of the above-mentioned problems of the prior art, and an object of the present invention is to provide a neural network having an efficient evaluation value in trials with different initial values and conditions. Is to provide a neural network that can perform learning efficiently.

[0010]

Means for Solving the Problems The above-mentioned problems are solved in the present invention as follows. (1) A neural network group is formed by arranging two kinds of multilayer hierarchical neural networks in parallel, one of which is a traveling neural network and the other is a comparing neural network. The two types of neural networks are set independently, and the learning process of the neural network is applied in parallel. The neural network learning process is applied up to
The comparison side neural network resets the initial value newly, applies the learning processing of the neural network up to the number of segmentation learning, and repeats the above-described processing. Conversely, when the comparison side neural network has a good evaluation value, If the number of times and the number of times of learning on the comparison side are not the same, the learning processing of the neural network is applied to the neural network on the side of comparison until the number of times of learning becomes the same as that of the neural network on the progress side. Is set in the progress-side neural network, the comparison-side neural network resets the initial value, and repeats the above-described process by applying the neural network's learning process up to the number of segment learning. (2) In the above (1), initial values and conditions of each neural network, that is, each coupling coefficient, learning method, learning coefficient, number of layers, number of hidden units, and the like are set by random numbers, previously created procedures, input devices, and the like. I do.

Next, the operation will be described. In the invention of claims 1 and 2 of the present invention, as described in the above (1) and (2), when applying the learning process of the neural network, two types of neural networks are arranged in parallel, and each neural network is assigned to each neural network. Since neural network learning processing is applied and evaluation is performed for each number of times of segmentation learning, trials of neural networks with bad initial values and conditions can be terminated earlier, and conversely neural networks with good initial values and conditions can be A neural network with a good evaluation value can be obtained by a neural network learning processing method capable of selecting and proceeding with learning. In addition, initial values and conditions for each neural network, that is, each coupling coefficient, a learning method, a learning coefficient, the number of layers, the number of hidden units, and the like can be set.
Various trials can be compared and selected. For this reason, the neural network with good conditions can learn more, can learn efficiently, and can shorten the calculation time.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the neural network according to the present invention will be described below in detail with reference to the drawings. (1) In claim 1, two types of multilayer hierarchical neural networks are arranged in parallel to form a neural network group, one of which is a traveling-side neural network and the other is a comparison-side neural network. . The two types of neural networks are set independently, and the learning process of the neural network is applied in parallel. The learning process of the neural network is applied up to the number of departure learning times, the initial value of the comparison side neural network is newly reset, the learning process of the neural network is applied up to the number of departure learning times, and the above process is repeated. If the neural network has a good evaluation value and the number of learnings on the traveling side and the number of learnings on the comparison side are not the same, the learning processing of the neural network is applied to the comparing side neural network until the number of learnings on the traveling side becomes equal to the number of learnings on the comparing side If they are the same, All values of the neural network are set to the progressing neural network, the comparison neural network resets the initial value newly, applies the neural network learning process up to the number of segmentation learning, and repeats the above process An embodiment of the network is shown by the flowcharts of FIGS. 1 to 6 and 7 to 9. The detailed procedure of FIGS. 1 to 6 and FIGS. 7 to 9 will be described later.

(2) In claim 2, the initial values and conditions of each neural network, that is, each coupling coefficient, learning method, learning coefficient, number of layers, number of hidden units, and the like are determined by random numbers, previously created procedures, input devices, and the like. Embodiments of the neural network to be set are shown by flowcharts in FIGS. 1 to 6 and FIGS. 7 to 9. 1 to 6
The detailed procedure of FIGS. 7 to 9 will be described later. FIG. 1 is an overall configuration diagram of an embodiment of a learning / recognition device including a neural network according to the present invention. Reference numeral 10 denotes a neural network learning / recognition device, 1 denotes a neural network control unit, and 2 denotes a working memory (neural network control unit 1). 3 is an input control unit, 4 is an output control unit, 5 is an input device, 6 is an output / display device, and 7 is a condition setting table. , 8 are a group of neural networks.

In the neural network learning / recognition device 10, when various set values such as a learning pattern and a teacher pattern are input to the input device 5, the input values are stored in the working memory 2 via the input control unit 3. It is memorized. Each input value is input to the neural network 1, where learning is performed by the neural network control unit 1 based on the learning pattern on the working memory 2, the teacher pattern, and various setting values in the condition table 7 to determine the coupling coefficient. Update. The coupling coefficient is stored in the neural network group 8 and the learning result is output and displayed on the output / display device 6 via the output control unit 4. In the neural network learning / recognition device 10, the unknown pattern is input to the input device 5.
When input is made, the input value is stored in the working memory 2 via the input control unit 3 as described above.
The neural network control unit 1 performs recognition or value estimation based on the coupling coefficient in the neural network group 8 and outputs / displays the result to the output / display unit 6 via the output control unit 4.

Next, an embodiment of the present invention will be described with reference to FIGS. FIG. 2 is a conceptual diagram showing a connection state by a unit of a neural network composed of input, intermediate and output layers, and FIG. 3 is an explanatory diagram of an internal state of one unit. FIG. 4 shows an embodiment of a neural network for performing digit recognition, in which a feature of numeric data is given as an input pattern.
This is a system that determines which number is based on the output pattern. In this neural network, the number of layers is three, the unit of the input layer ６４ is 64 units representing the characteristic amount of the character, the unit of the intermediate layer ８ is 8 units connected to the unit of each input layer by a coupling coefficient, and the output layer is 〓 is composed of 10 units each representing a number from 1 to 9 and 0. In this example, the teacher pattern is 10000000 in the case of the numeral 1, for example.
00 pattern is given. This means that the first one indicates that the first value of the output unit is one,
The following output unit values indicate 0. For other numbers, a teacher pattern is created similarly. Each output unit uses this value at the time of learning and recognition determination. When selecting the optimum unit at the time of judgment, the calculated value of the unit of the output layer is, for example, 0.1, 0.9, 0.3
, ..., 0.1, 0.2 If the second value of the output unit is
Since it is 0.9, which is the maximum, it is judged to be the number 2.

FIG. 5 is a diagram showing an example of the interior of the condition setting table 7 shown in FIG. 1 and shows the number of final learnings and the number of segmented learnings. Here, the number of times of segment learning means that the number of times of learning in the learning process of the neural network is processed by this number of times, and after this number of times of processing, the evaluation value is compared between the progressing neural network and the comparison side neural network.

FIG. 6 is a diagram showing an example of the evaluation values of the traveling-side neural network and the comparison-side neural network for each segment learning number in the evaluation value table. It can be displayed and output as needed. The figure shows the evaluation values of the traveling-side neural network and the comparison-side neural network with respect to each segment learning count. In this case, the evaluation value of the traveling-side neural network is 0.076 when the number of segment learning is 1,000, the evaluation value of the comparison-side neural network is 0.085, and the evaluation value of the traveling neural network is 0.0 when the number of segment learning is 2,000. 054, the evaluation value of the comparison side neural network has not been obtained. The evaluation value of the progress side neural network is 0.032 at the delimitation learning number of 3,000. The evaluation value of the comparison side neural network has not been obtained. 000 indicates that the evaluation values of the traveling-side neural network and the comparison-side neural network have not been determined. Here, the evaluation value of the neural network is the value of the loss function of the learning result of the neural network. Thus, less indicates better.

FIGS. 7 and 8 are flowcharts for explaining the learning means of the neural network. FIG. 9 is a flowchart for explaining the recognition means of the neural network. Hereinafter, a procedure for performing learning using the flowcharts of FIGS. 7 and 8 will be described. 11 ... Process start. 12 ... In the neural network control unit 1,
The number of times of learning of the traveling-side neural network in the neural network group 8 is set to zero. In the flowchart, the neural network is abbreviated as NN.
The value input from the input device 5 is passed through the input control unit 3 to the neural network control unit 1 or the working memory 2
Set the value of each parameter in. The values obtained by inputting the number of final learning and the number of segment learning in the condition table 7 from the input device 5 are set via the input control unit 3. 13 ... In the neural network control unit 1,
A coupling coefficient and the like in the traveling-side neural network are initialized by random numbers and the like. Also, the learning method, learning coefficient, number of layers,
The number of units in the intermediate layer and the like are set by a random number, a procedure created in advance, or the input device 5. 14 In the neural network control unit 1,
Initialize the coupling coefficients and the like in the comparison-side neural network with random numbers and the like. The number of times of learning of the comparison side neural network is set to zero. Further, the learning method, the learning coefficient, the number of layers, the number of hidden units, and the like are set by random numbers or by a previously created procedure or by the input device 5 or the like.

15. In the neural network controller 1, learning of the neural network is applied to the traveling-side neural network by the number of times of the partition learning. The coupling coefficient is held in the traveling neural network in the neural network group 8. This is shown in FIG.
This will be described with reference to the flowchart of FIG. 16 ... Is the calculation end condition satisfied? If yes, go to 19. If no, go to the next step 17. Here, the calculation termination condition is one specified processing time. 2 The specified number of repetitions has been reached. 3 The value of the loss function has fallen below the specified value. Four
The specified recognition rate has been reached. 5 Below the false recognition rate. It was less than 6 designated unrecognition rates. 7 The classification of learning data and recognition data no longer changes. Etc. and its composite conditions. Here, the recognition rate, the false recognition rate, and the non-recognition rate are the ratios of the number of verification data and the number of correct recognition data, the number of false recognition data, and the number of non-recognition data used for the determination. 17 ... In the neural network control unit 1,
The learning of the neural network is applied to the comparison side neural network by the number of times of the partition learning. The coupling coefficient is held in the comparison side neural network in the neural network group 8. This will be described in detail with reference to the flowchart of FIG. 18 ... Is the calculation end condition satisfied? If yes, go to 19. If NO, go to the next 21. 19: In the neural network control unit 1,
Output control unit 4 outputs a result such as a coupling coefficient obtained by learning.
Is stored and output and displayed in the device 6 that performs output and display through the. 20 ... The learning process ends. 21: In the neural network control unit 1,
The latest neural network evaluation value and the evaluation value of the progressive neural network having the same number of learning times as the comparison neural network are set for comparison preparation. The evaluation value is a value of the loss function of the learning result of the neural network. 22... In the neural network control unit 1
It is compared whether the evaluation value of the traveling-side neural network is better, that is, smaller than the evaluation value of the comparison-side neural network. In the case of YES, the process returns to 14 and the process is repeated.
If NO, go to the next step 23. 23: It is compared whether the number of times of learning of the traveling side neural network is equal to the number of times of learning of the comparison side neural network. If yes, go to 24. 17 in case of NO
And repeat the process. 24 In the neural network control unit 1,
Each value of the comparison side neural network is set to the progress side neural network. Returning to step 14, the process is repeated.

101: Processing started. 102. In the neural network control unit 1,
A learning pattern number in a predetermined neural network in the neural network group 8 is initialized to zero. Initialize the repeat count to zero. Set the repetition end count. Set the value of each unit number of the input layer, the middle layer, and the output layer. All learning patterns and teacher values are input from the input device 5 and set in the working memory 2 via the input control unit 3. As the coupling coefficient, a value held in a predetermined neural network in the neural network group 8 is used. The number of learning patterns is input from the input device 5 and set. 103. In the neural network control unit 1,
A learning pattern is set for each unit in the input layer.

104. The neural network controller 1 calculates the output value of the intermediate layer unit based on each unit of the input layer and the coupling coefficient between the input layer and the intermediate layer. The output value of the output layer unit is calculated from each unit of the intermediate layer and the coupling coefficient between the intermediate layer and the output layer. 105. In the neural network control unit 1,
An error value between the teacher value and the output value of the output layer unit is calculated. Further, the error value of the output layer unit is propagated to the intermediate layer side to calculate the error value of the intermediate layer. 106. In the neural network control unit 1,
The coupling coefficient between the intermediate layer and the output layer is updated based on the error value. Similarly, the coupling coefficient between the input layer and the intermediate layer is updated. 107. In the neural network control unit 1,
Count up the learning pattern number. 108. In the neural network control unit 1,
It is determined whether the learning pattern is completed. If yes, go to
If NO, go to 105. 109. In the neural network control unit 1,
It is determined whether the sum of the error values is equal to or less than a specified value. If YES, go to 113; if NO, go to next.

110. The neural network controller 1 determines whether all output patterns and the teacher pattern match. If YES, go to 113; if NO, go to next. 111. In the neural network control unit 1,
Count up the repetition count. The learning pattern number is re-initialized to 0. 112. In the neural network control unit 1,
It is determined whether the repetition is completed. If YES, go to next, NO
If so, go to 103. 113. In the neural network control unit 1,
The evaluation value is stored at a predetermined number of breaks in the evaluation value table of FIG. 6 and at a position on the corresponding matrix of the neural network. End. Proceed to the position where 101 was called.

Next, a procedure for performing recognition using the flowchart of FIG. 9 will be described. 201 ... Processing started. 202. In the neural network control unit 1,
The neural network model name is input from the input device 5 via the input control unit 3 and the neural network group 8
Is set in a predetermined neural network. Read the number of each unit of the input layer, the middle layer, and the output layer and set the value. Read and set the values of various parameters. Read and set the learned coupling coefficients. 203. Input recognition data or unknown data from the input device 5. 204. In the neural network control unit 1,
The recognition data or unknown data is set in the input layer, and the output values of the intermediate layer and the output layer unit are calculated based on the respective coupling coefficient values. 205. In the neural network control unit 1,
The judgment result is calculated from the output value of the output layer unit and the judgment condition. 206. In the neural network control unit 1,
The determination result is output and displayed on the device 6 for outputting and displaying the result through the output control unit 4. 207. In the neural network control unit 1,
Is recognition data or unknown data end? If YES, go to the next step. If NO, go to 203. 208 ... End. Proceed to the place where 201 was called.

[0024]

As described above, according to the neural network of the present invention, two types of neural networks are arranged in parallel for learning, recognition and estimation of numeral recognition, character recognition, shape recognition, and the like. Since the learning process is applied to the network and evaluation is performed for each number of times of learning, the trial of the neural network with bad initial values and conditions is terminated early, and the neural network with good initial values and conditions is selected and advanced. A neural network with a good evaluation value can be obtained by a network technique. Further, a learning method, a learning coefficient, the number of layers, the number of hidden units, and the like can be set for each neural network, and various trials can be compared and selected. For this reason, a neural network with good conditions has more advanced learning, can obtain good evaluation values efficiently, and can reduce the calculation time, and is extremely effective in practice.

[Brief description of the drawings]

FIG. 1 is a system configuration diagram of an embodiment employing a neural network according to the present invention.

FIG. 2 is a schematic diagram of unit connection of a three-layered neural network.

FIG. 3 is an explanatory diagram of a unit internal state.

FIG. 4 is an explanatory diagram of one embodiment of a neural network for performing digit recognition.

FIG. 5 is a diagram showing a final learning count and a break learning count in a condition table according to an embodiment of the present invention.

FIG. 6 is a diagram showing evaluation values of a traveling-side neural network and a comparison-side neural network according to one embodiment of the present invention.

FIG. 7 is a flowchart illustrating learning means of the neural network.

FIG. 8 is a flowchart illustrating learning means of a neural network.

FIG. 9 is a flowchart for explaining a neural network recognition and estimation unit.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Neural network control part 2 Working memory 3 Input control part 4 Output control part 5 Input device 6 Device which performs output and display 7 Condition setting table 8 Neural network group 10 Neural network learning and recognition device

Claims (2)

[Claims] A neural network group is formed by arranging two types of multilayer hierarchical neural networks in parallel, one of which is a traveling-side neural network and the other is a comparing-side neural network. The neural network is independently set, performs a learning process of the neural network in parallel, and is a learning method of a neural network that performs a learning process until a desired condition is satisfied.
Set a mechanism to compare the learning results of the neural network for each learning count, advance the selection, replacement, and learning, compare the learning results of the neural network for each partition learning count, and select a neural network with a good evaluation value Replacing the traveling side neural network,
Other neural networks are characterized by resetting a new initial value, learning the neural network up to the number of times of segmentation learning, and repeating the processing described above. 2. The initial values and conditions of each neural network, ie, each coupling coefficient, learning method, learning coefficient, number of layers,
The number of units in the middle layer, etc.
2. The neural network according to claim 1, wherein the neural network is set by an input device or the like.