JP3631443B2 - Neural network connection weight optimization apparatus, method and program - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a neural network connection weight optimization apparatus and method for optimizing the connection weight of a neural network based on a model evaluation function that determines the preference of the neural network.
[0002]
[Prior art]
A neural network is a computer technology developed to artificially realize advanced information processing functions. In other words, a neural network is an artificial intelligence technology that uses a computer to simulate a neural circuit in the brain. The input layer is a collection of input elements for inputting information. It consists of an output layer, which is a collection of output elements that output whether or not, and an intermediate layer, which is a collection of intermediate elements in the middle of those layers. They are connected.
[0003]
A neural network acquires advanced information processing functions by optimizing the connection weight parameter that takes a high evaluation function value (or a low evaluation function value depending on the evaluation function) based on some model evaluation function. Can do.
[0004]
As a method for optimizing the connection weight parameter of the neural network, an error back-propagation method (for example, Rumelhalt, DE, GE Hinton, and RJ Williams, “Learning representations by back-propagating errors. ”, Nature 323, pp. 533-536., 1986.), but there is a problem that many constraints are required for the model evaluation function, such as partial differential calculation is possible for all parameters. .
[0005]
Further, as a parameter optimization method for a general model evaluation function, a genetic algorithm (for example, Goldberg, DE, “Genetic Algorithms in Search, Optimization, and Machining Learning.”, Addison-Wesley Publishing). Company Inc., 1989.), however, there is a problem that sufficient optimization performance cannot be obtained because the connection weight values of the neural network are simply collected and used as a real vector.
[0006]
[Problems to be solved by the invention]
As a conventional technique for acquiring advanced information processing functions by optimizing the neural network connection weight parameters with respect to the model evaluation function, we specialize in neural network connection weight optimization such as the error back propagation method. There is a problem that the method cannot be used for a general model evaluation function, and a method such as a genetic algorithm using a real vector that collects connection weights of a neural network uses a general model evaluation function. However, there was a problem that sufficient optimization performance could not be obtained.
[0007]
The present invention has been made in consideration of the above circumstances, and is a neural network connection weight optimization device that enables optimization better than conventional optimization of connection weights of a neural network based on a general model evaluation function. And to provide a method.
[0008]
[Means for Solving the Problems]
The present invention is a neural network connection weight optimization method for optimizing a connection weight of a neural network having a predetermined connection structure composed of a plurality of network elements and zero or one or more bias elements. A first step of associating a neural network with an evaluation result obtained by evaluating a connection weight of the neural network based on a predetermined evaluation function, and storing it in a solution candidate neural network set storage means; and the solution candidate neural network set storage For each of a predetermined number of neural networks selected from the neural network stored in the means, the function of the network element is considered for each connection weight related to the input to the same network element among all the connection weights of the neural network. Encode by performing a predetermined encoding function conversion, for each network element A second step of generating a coded neural network by concatenating the generated codes and a new number of coded neural networks based on the coded neural network generated in the second step. For each of the coded neural nets generated in the third step, and for each code part corresponding to each network element among all the codes of the coded neural network, a network is generated. A fourth step of newly generating a neural network of the same format as the neural network stored in the solution candidate neural network set storage means by performing a predetermined inverse function transformation considering the function of the element; Each of the neural networks generated in step 4 is evaluated based on the model evaluation function. Determining whether to add the neural network and its evaluation result to the solution candidate neural network set storage means based on at least the evaluation result, and the neural network determined to be added and its evaluation result And a fifth step of storing them in the solution candidate neural network set storage means in association with each other.
[0009]
Preferably, in the fifth step, among the existing neural nets stored in the solution candidate neural net set storage means and evaluation results thereof, those satisfying a predetermined condition are stored in the solution candidate neural network set storage. You may make it delete from a means.
[0010]
Preferably, the second step to the fifth step are repeatedly executed until it is determined that a predetermined end condition is satisfied after the fifth step, and the predetermined step is performed after the fifth step. If it is determined that the termination condition is satisfied, among the neural networks stored in the solution candidate neural network set storage means, those having the best evaluation result at that time are optimized. You may make it select as a neural network.
[0011]
Preferably, in the second step, when the predetermined encoding function conversion is performed on the connection weights related to the input to the same network element of the neural network, all the inputs related to the input to the network element are performed. Obtain a vector length of a coupling weight vector composed of coupling weights other than the coupling weights related to the input from the bias element among the coupling weights, and input to the network element based on the vector length or a value obtained by multiplying the vector length by a positive number. The coded neural network is generated by concatenating the value obtained by performing a predetermined operation on each of all the connection weights according to the above and the vector length or a value obtained by multiplying the vector length by a positive number as a code. You may do it. Preferably, the predetermined operation may be an operation of dividing the coupling weight by the vector length or a value obtained by multiplying the vector length by a positive number.
[0012]
Preferably, in the fourth step, when the predetermined inverse function transformation is performed on a code portion corresponding to one network element among all codes of the coding neural network, Based on a code value corresponding to a vector length or a value obtained by multiplying the vector length by a positive number, a code value corresponding to a value obtained by performing a predetermined operation on each of the coupling weights in the code portion, respectively. A corresponding connection weight in the new neural network may be obtained by performing a predetermined calculation. Preferably, the predetermined calculation obtains a correction value obtained by performing a predetermined correction on a code value corresponding to the vector length or a value obtained by multiplying the vector length by a positive number in the code portion, and Of these, a code value corresponding to a value obtained by performing a predetermined calculation on each of the coupling weights may be calculated by multiplying the correction value by each. Preferably, the predetermined correction is a correction that takes an absolute value for the value, or if the value is less than a predetermined positive threshold, the value is the predetermined positive number. You may make it the correction used as a threshold value.
[0013]
Preferably, in the third step, the encoded neural network may be generated by a genetic algorithm.
[0014]
The present invention relating to the apparatus is also established as an invention relating to a method, and the present invention relating to a method is also established as an invention relating to an apparatus.
Further, the present invention relating to an apparatus or a method has a function for causing a computer to execute a procedure corresponding to the invention (or for causing a computer to function as a means corresponding to the invention, or for a computer to have a function corresponding to the invention. It is also established as a program (for realizing) and also as a computer-readable recording medium on which the program is recorded.
[0015]
Now, the function of the neural network is realized by collecting the functions of each network element, and it can be considered that only the connection weight input to the element influences the function of each network element. In the present invention, all the connection weights of a neural network are grouped for each subset of connection weights input to each network element, and an encoded neural network is generated by performing an encoding function conversion to generate an encoded neural network. A new neural network set is generated by performing inverse encoding function conversion for each subset of connection weights input to each network element. As an optimization of the connection weight of the neural network based on the model evaluation function, it is possible to perform an optimization superior to the conventional one. Further, optimization can be performed for each function of each network element. Further, according to the present invention, a neural network element that is functionally similar but has a significantly different weight value can be converted into a similar code vector. Optimization can be performed, and high-performance optimization can be realized.
[0016]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the invention will be described with reference to the drawings.
[0017]
FIG. 1 shows a configuration example of a neural network connection weight optimizing device that performs an optimization process related to a connection weight parameter of a neural network according to an embodiment of the present invention.
[0018]
As shown in FIG. 1, the neural network connection weight optimization apparatus according to the present embodiment includes a solution candidate neural network set storage unit 101, a model evaluation function storage unit 102, a neural network element encoding unit 103, an encoded neural network set. A storage unit 104, a new encoded neural network generation unit 105, a new encoded neural network set storage unit 106, a neural network element decoding unit 107, a new neural network set storage unit 108, and a new neural network evaluation / selection unit 109 are provided. Yes.
[0019]
This neural network connection weight optimization device can be realized by software (that is, can be realized by executing a program on a computer). At that time, a part or all of the functions of the software can be realized as a chip or board and incorporated in the computer. In addition, when the neural network connection weight optimization apparatus is realized by software, it can be incorporated as a function of other software. It is also possible to configure this neural network connection weight optimization device as dedicated hardware.
[0020]
The solution candidate neural network set storage unit 101, the model evaluation function storage unit 102, the encoded neural network set storage unit 104, the new encoded neural network set storage unit 106, and the new neural network set storage unit 108 are all, for example, a hard disk, It is configured by a storage device such as an optical disk or a semiconductor memory. Each storage unit may be configured by a separate storage device, or all or a part of them may be configured by the same storage device.
[0021]
Although not shown in FIG. 1, the neural network connection weight optimization device includes an input / output device for exchanging data with the outside. Of course, a GUI (graphical user interface) may be provided, or a network connection interface may be provided.
[0022]
FIG. 2 shows an example of the overall processing procedure of the present neural network connection weight optimization apparatus.
[0023]
The model evaluation function storage unit 102 stores data indicating a model evaluation function for evaluating the neural network in advance. A desired general model evaluation function can be used as the model evaluation function.
[0024]
The solution candidate neural network set storage unit 101 stores at least one piece of data indicating an initial neural network created in advance. This initial neural network generation method may be in accordance with a conventional method (for example, the structure of the neural network is created or selected by the user, and the connection weight is randomly generated by a predetermined algorithm). Each of the neural networks stored in advance in the solution candidate neural network set storage unit 101 is evaluated by the same method as the evaluation performed on the new neural network in step S4, and the evaluation result (for example, evaluation It is assumed that (function value) is stored in association with each neural network.
[0025]
First, the neural network element encoding unit 103 extracts n (n is a predetermined or designated integer greater than or equal to 1) neural network from the solution candidate neural network set storage unit 101, and the extracted n neural networks. For each of the nets, n coding neural nets are generated by treating the connection weights input to the same element in the neural network as the same group and performing a predetermined coding function conversion, and these are coded neural networks. The data is stored in the net set storage unit 104 (step S1).
[0026]
Next, the new encoded neural network generation unit 105 generates a predetermined method such as a genetic algorithm based on the encoded neural network generated in step S1 and stored in the encoded neural network set storage unit 104. Accordingly, m new encoded neural nets (m is a predetermined or designated integer or more) are generated and stored in the new encoded neural network set storage unit 106 (step S2).
[0027]
As this generation method, various methods can be used. For example, the blx method (for example, Eshelman, LJ, and Schaffer, JD, “Real-Coded Genetic Algorithms and Interval-Schemata.”, Foundations of Genetic Algorithms. .) May be used.
[0028]
Next, the neural network element decoding unit 107 extracts the m new encoded neural nets generated in step S2 and stored in the new encoded neural network set storage unit 106, and the m neural nets extracted are extracted. For each, the codes corresponding to the same elements in the neural network are treated as the same group, and predetermined inverse function transformation is performed to generate m new neural networks, which are stored in the new neural network set storage unit 108. Store (step S3).
[0029]
Next, the new neural network evaluation / selection unit 109 obtains the information stored in the model evaluation function storage unit 102 for each of the m new neural networks stored in the new neural network set storage unit 108 in step S3. Based on the evaluation result, all or part of the solution candidate neural nets stored in the solution candidate neural network set storage unit 101 and m stored in the new neural network set storage unit 108 are evaluated. One or more neural nets are selected from the new neural nets (or from the m new neural nets stored in the new neural net set storage unit 108), and data of the selected new neural nets are selected. For this, the solution candidate neural network is used as a new solution candidate neural network together with the evaluation result. Stored in the case memory portion 101 (step S4). When storing the selected new neural network in the solution candidate neural network set storage unit 101, one or more solution candidate neural networks selected on the basis of a predetermined criterion are deleted from the solution candidate neural network set storage unit 101. May be.
[0030]
As this evaluation / selection method, various methods can be used. For example, the mgg method (for example, Sato, Ono, Kobayashi, “Proposal and Evaluation of Generation Alternation Model in Genetic Algorithm”, Japanese Society for Artificial Intelligence, Vol. 12, No. 5, pp. 734-744., 1997.) May be used.
[0031]
Here, it is determined whether or not a predetermined end condition set in advance is satisfied (step S5). If not satisfied, the process returns to step S1 and the same processing is repeated.
[0032]
As the end condition, for example, the number of repetitions reaches a predetermined number, the difference between the number of times and the previous evaluation result (for example, the evaluation function value) falls below a predetermined value, Various things can be considered, such as the rate of change between the first and previous evaluation results (e.g., evaluation function value) being lower than a predetermined value, and a combination of these conditions as a logical sum or logical product.
[0033]
When the termination condition is satisfied, the processing loop is exited and the neural network stored in the solution candidate neural network set storage unit 101 is selected with the best evaluation result. This selected neural network is a neural network obtained by this optimization processing.
[0034]
Here, when a genetic algorithm is used in the procedure of FIG. 2, one loop process is as shown in FIG. 3, for example. That is, for example, two neural networks are randomly selected and extracted from the solution candidate neural network set storage unit 101 in step S1 of this time, and the respective neural nets a and b are converted into encoded neural nets. In step S3, for example, four new encoded neural nets are generated by applying a genetic algorithm to these encoded neural nets. Then, in step S3, they are decoded and inversely converted into new neural nets c to f. It is stored in the storage unit 108. Next, in step S4, the new neural nets c to f are evaluated, and then, for example, the best two of the original solution candidate neural net a and the new neural nets c to f are selected based on the evaluation results. If they are not the original neural network (a, b), for example, the original neural nets a, b are updated with the selected neural nets e, f. Of course, this is only an example, and various other methods are possible as one loop processing when a genetic algorithm is used, and various methods using an algorithm other than the genetic algorithm are possible. Is possible.
[0035]
Now, neural network element encoding and neural network element decoding in the neural network connection weight optimizing device of this embodiment will be described using specific examples.
[0036]
FIG. 4 shows the stored contents of the solution candidate neural network set storage unit 101, the processing of the neural network element encoding unit 103, and the stored content of the encoded neural network set storage unit 104 in the neural network connection weight optimization apparatus of FIG. This is just an example.
[0037]
As described above, the solution candidate neural network set storage unit 101 stores one or more solution candidate neural networks (201), each associated with an evaluation result. When a plurality of solution candidate neural networks are stored, all solution candidate neural networks have the same structure.
[0038]
The solution candidate neural network 201 illustrated in FIG. 4 is coupled to three network elements (input elements) i1 to i3 and one bias element to three network elements (intermediate elements) h1 to h3, respectively. In this example, intermediate elements h1 to h3 and one bias element are coupled to two network elements (output elements) o1 and u, respectively.
[0039]
In the neural network element encoding unit 103, for each solution candidate neural network, all connection weights of the solution candidate neural network are grouped for each connection weight input to the same network element, and an encoding function conversion (203) is performed. To generate a coded neural network. For example, the coding function transformation (203) is performed on the connection weight group (202) input to the network element u in the solution candidate neural network (201) of FIG. A portion (204) corresponding to is generated. Such a process is performed on all the network elements h1, h2, h3, o1, u and connected to generate an encoded neural network. The generated encoded neural network (205) is stored in the encoded neural network set storage unit 104.
[0040]
Here, the encoding function conversion will be described in more detail.
[0041]
First, FIG. 5 shows an encoded neural network generation method (1203) according to a conventional method. According to the conventional method, an encoded neural network is generated by arranging all connection weights input to all network elements in order.
[0042]
For example, the connection weight W input to the network element u (1202)_{u, 1}~ W_{u, b}, The portion corresponding to u of the encoded neural network (1204),
{W_{u, 1}  W_{u, 2}  W_{u, 3}  W_{u, b}}
Will be generated. The same applies to the other network elements h1, h2, h3, o1.
[0043]
And portions corresponding to the respective network elements h1, h2, h3, o1, u,
{W_h1,1  W_h1,2  W_h1,3  W_{h1, b}}
{W_h2,1  W_h2,2  W_h2,3  W_{h2, b}}
{W_h3,1  W_h3,2  W_h3,3  W_{h3, b}}
{W_o1,1  W_o1,2  W_o1,3  W_{o1, b}}
{W_{u, 1}  W_{u, 2}  W_{u, 3}  W_{u, b}}
By connecting the coding neural network,
{W_h1,1  W_h1,2  W_h1,3  W_{h1, b}  W_h2,1  ... W_{o1, b}  W_{u, 1}  W_{u, 2}  W_{u, 3}  W_{u, b}}
Is generated.
[0044]
Therefore, for example, in the solution candidate neural network (201) of FIG. , −2.0}. Similarly, the coupling weights input to the network element h2 are {2.3, −1.3, 10.0, 1.0}, and the coupling weights input to the network element h3 are , {2.1, 2.2, −3.3, −2.0}, and the coupling weights input from the network elements h1, h2, h3 and the bias element to the network element o1 are {−5 .1, 2.4, −1.0, 2.0}, and similarly, the connection weights input to the network element u are {−0.1, 20.4, −0.8, 1.0. } Is obtained by the conventional method. Encoding neural net,
{0.1, 2.3, 3.0, -2.0, 2.3, -1.3, 10.0, 1.0, 2.1, 2.2, -3.3, -2 0.0, -5.1, 2.4, -1.0, 2.0, -0.1, 20.4, -0.8, 1.0}
It becomes.
[0045]
Next, FIG. 6 shows an example of the encoding function conversion (203) in the present embodiment. In this example, for each network element, the vector length of the coupling weight vector other than the coupling weight with the bias element among the weights included in the subset of coupling weights input to the network element is calculated, and the coupling weight part is calculated. An encoded neural network corresponding to the network element is generated by concatenating all elements of the set divided by the vector length and the vector length as a code.
[0046]
For example, the connection weight W input to the network element u (202)_{u, 1}~ W_{u, b}W_{u, 1}~ W_{u, 3}Vector length L when_u,
L_u= (W_{u, 1} ²+ W_{u, 2} ²+ W_{u, 3} ²)^1/2
And W_{u, 1}~ W_{u, b}Each L_uDivide by
w_{u, 1}= W_{u, 1}/ L_u
w_{u, 2}= W_{u, 2}/ L_u
w_{u, 3}= W_{u, 3}/ L_u
w_{u, b}= W_{u, b}/ L_u
And w_{u, 1}, W_{u, 2}, W_{u, 3}, W_{u, b}, L_u, The portion corresponding to u of the encoded neural network (204),
{W_{u, 1}  w_{u, 2}  w_{u, 3}  w_{u, b}  L_u}
Is generated. The same applies to the other network elements h1, h2, h3, o1.
[0047]
And portions corresponding to the respective network elements h1, h2, h3, o1, u,
{W_h1,1  w_h1,2  w_h1,3  w_{h1, b}  L_h1}
{W_h2,1  w_h2,2  w_h2,3  w_{h2, b}  L_h2}
{W_h3,1  w_h3,2  w_h3,3  w_{h3, b}  L_h3}
{W_o1,1  w_o1,2  w_o1,3  w_{o1, b}  L_o1}
{W_{u, 1}  w_{u, 2}  w_{u, 3}  w_{u, b}  L_u}
By connecting the coding neural network,
{W_h1,1  w_h1,2  w_h1,3  w_{h1, b}  L_h1  w_h2,1  … W_{o1, b}  L_o1  w_{u, 1}  w_{u, 2}  w_{u, 3}  w_{u, b}  L_u}
Is generated.
[0048]
Therefore, for example, if the solution candidate neural network (201) in FIG. 4 has the coupling weights exemplified above, the encoding neural network obtained by the encoding function conversion according to this example is
{0.0264442 ..., 0.6082187 ..., 0.7933288 ..., -0.5288858 ..., 3.7815340 ...,
0.2223370 ..., -0.1256874 ..., 0.9668269 ..., 0.0966826 ..., 10.343113 ...,
0.4679393 ..., 0.4902221 ..., -0.7353332 ..., -0.4456565 ..., 4.4877611 ...,
-0.8909061 ..., 0.4192499 ..., -0.1746874 ..., 0.349373749 ..., 5.72445087 ...,
-0.0048981 ..., 0.9992199 ..., -0.0391850 ..., 0.0489813 ..., 20.4159925 ...}
It becomes.
[0049]
In this way, neural network elements that are functionally similar but have significantly different weights can be converted into similar code vectors (eg, {10, 20, 30, -40} and {20, 40, 60, −80} can be converted into {1, 2, 3, −4, 10} and {1, 2, 3, −4, 20} vectors), on the function of each network element It is possible to optimize the connection weight parameter that avoids duplication, and to realize high-performance optimization.
[0050]
Next, FIG. 7 illustrates the processing of the neural network decoding unit 107 in the neural network connection weight optimization apparatus of FIG.
[0051]
As described above, the neural network element decoding unit 107 generates a new neural network by treating codes corresponding to the same element in the new encoded neural network as the same group and performing a predetermined inverse function conversion.
[0052]
Here, the inverse function conversion will be described in more detail.
[0053]
In the example of FIG. 7, for each of the code vectors corresponding to each network element in the new encoding neural network, the code vector is obtained by multiplying the other variables by the absolute value of the variable representing the length of the code vector. A new neural network part corresponding to is generated.
[0054]
For example, the code (301) corresponding to the element u in the new encoding neural network,
{W_{u, 1} ^new  w_{u, 2} ^new  w_{u, 3} ^new  w_{u, b} ^new  L_u ^new}
For, the code L corresponding to the length of the vector_u ^new(L_u ^newMay be negative) | L_u ^new｜ L_u ^newW corresponding to each combination weight other than_{u, 1} ^new~ W_{u, b} ^newEach | L_u ^newMultiply
W_{u, 1} ^new  = W_{u, 1} ^new  ・ | L_u ^new｜
W_{u, 2} ^new  = W_{u, 2} ^new  ・ | L_u ^new｜
W_{u, 3} ^new  = W_{u, 3} ^new  ・ | L_u ^new｜
W_{u, b} ^new  = W_{u, b} ^new  ・ | L_u ^new｜
And the connection weight (303) to be input to the element u of the new neural network
{W_{u, 1} ^new  W_{u, 2} ^new  W_{u, 3} ^new  W_{u, b} ^new}
Ask for. The same applies to the codes corresponding to the elements h1, h2, h3 and o1 in the new encoded neural network.
[0055]
Various variations can be considered for the inverse function transformation.
[0056]
For example, in the above, L^newInstead of L, but instead of L^newIs a predetermined threshold value L_thIf this is the case, the L^newAnd L^newIs a predetermined threshold value L_thL if less than^new= L_thThere is also a method.
[0057]
Also, in the above, L^new2 may be negative, but instead, in step S2 of FIG.^newAdopt an algorithm that does not become negative or L^newEven if the algorithm may be negative (or less than 0)^newL becomes negative (or 0 or less)^newMay be corrected to 0 or more (or a positive value), or the new coded neural network may not be used (discarded). In this case, L^newProcessing such as taking the absolute value of becomes unnecessary.
[0058]
In addition to the above, various variations can be considered for the encoding function conversion and the inverse function conversion.
[0059]
For example, in the coding function conversion example described above, the weight (W) included in the subset of connection weights input to the network element.₁~ W_n) And a coupling weight other than the coupling weight with the bias element (W₁~ W_tVector length L (= (W₁ ²+ ... + W_t ²)^1/2) And calculate all elements (W₁~ W_n) Divided by the vector length L and the vector length are connected as a code, so that the portion of the encoded neural network corresponding to the network element (W₁/L,...,W_n/ L, L), but instead, vector length L is multiplied by k (where k> 0) and all elements W₁~ W_nBy dividing L · k by L · k as a code, the portion of the encoded neural network corresponding to the network element (W₁/(L·k),...,W_nA method of generating / (L · k), (L · k)) is also possible. The inverse function transformation corresponding to this variation is the same as the example described so far if L · k is treated as L.
[0060]
Also, for example, the vector length L is multiplied by k (where k> 0), and the element W₁~ W_tDivided by L and element W_{t + 1}~ W_nIs divided by L · k, and L and L · k (or L and k) are connected as codes, thereby the portion of the encoded neural network corresponding to the network element (W₁/L,...,W_t/ L, W_{t + 1}/(L·k),...,W_n/ (L · k), L, (L · k)) (the vector length portion may be L or k instead of L or (L · k)). The inverse function transformation corresponding to this variation₁~ W_tL, L · k for W_{t + 1}~ W_nIt is the same as the example described so far if it is treated as L for.
[0061]
Further, for example, the vector length L is multiplied by k1 (where k1> 0), the vector length L is further multiplied by k2 (where k2> 0), and the element W₁~ W_tDivided by L · k1 and element W_{t + 1}~ W_nIs divided by L · k2, and L · k1 and L · k2 (or L, k1 and k2) are connected as a code to obtain a part of the encoded neural network corresponding to the network element (W₁/(L·k1),...,W_t/ (L · k1), W_{t + 1}/ (L · k2) ..., W_n/ (L · k2), (L · k1), (L · k2)) (the vector length part is L, k1, k2 instead of (L · k1), (L · k2)) Good) is also possible. The inverse function transformation corresponding to this variation uses L · k1 as W₁~ W_tL, L · k2 for W_{t + 1}~ W_nIt is the same as the example described so far if it is treated as L for.
[0062]
There are a method of fixing k, k1, and k2 and a method of determining each neural network.
[0063]
Of course, in addition to these, various variations can be considered for the encoding function conversion and the inverse function conversion.
[0064]
As described above, according to the present embodiment, it is possible to optimize the connection weight of the neural network for a general model evaluation function. Further, optimization can be performed for each function of each network element.
[0065]
In the following, the operation and effect of the connection weight optimization processing of the neural network of this embodiment will be described according to the following example.
[0066]
FIG. 8 illustrates a neural network having 10 input elements, 5 intermediate elements, and 1 output element, and optimizes the connection weight vector W (number of connection weights = 61) of this network.
[0067]
FIG. 9 shows an example of data (input and teacher signal) used for the model evaluation function f serving as the next optimization criterion.
f (W) = Σ (t (i) −O (in (i), W))²
Here, f (W): network model evaluation function value of the connection weight W
O (in (i), W): An output value of the network when in (i) is input. Moreover, the range which takes a sum total is i = 1-N.
In this example, a smaller f (W) network is evaluated as a preferable network.
When there are a plurality of output elements as in the example of FIG. 4, for example, the sum of model evaluation function values f (W) for each output element may be taken.
[0068]
FIGS. 10 and 11 show this embodiment in the case where the model evaluation function of FIG. 9 is used, the blx method is used in the new encoded neural network generation unit 105, and the mgg method is used in the new neural network evaluation / selection unit 109. 6 and 7 show the optimization results (C1, C3) and the optimization results (C2, C4) of the conventional method of FIG.
[0069]
The horizontal axis of FIGS. 10 and 11 is the number of iterations of the loop processing of the procedure of FIG. 2, the vertical axis of FIG. 10 indicates the change in the model evaluation function value, and the vertical axis of FIG. Shows the change in classification accuracy.
[0070]
From FIG. 10 and FIG. 11, it is clear that the optimization result of this embodiment is superior.
[0071]
As described above, according to the present embodiment, the neural network total connection weight is divided into groups that are input to the same element, and the coding function conversion is performed on each group, so that each element of the neural network can be independently performed. Since optimization can be performed, it is possible to optimize the coupling weight more efficiently than in the conventional method.
[0072]
Each function described above can be realized as software.
The present embodiment can also be implemented as a program for causing a computer to execute predetermined means (or for causing a computer to function as predetermined means, or for causing a computer to realize predetermined functions) The present invention can also be implemented as a computer-readable recording medium that records the program.
[0073]
Note that the configuration illustrated in the embodiment of the present invention is an example, and is not intended to exclude other configurations, and a part of the illustrated configuration may be replaced with another, or one of the illustrated configurations. Other configurations obtained by omitting a part, adding another function or element to the illustrated configuration, or combining them are also possible. Also, another configuration that is logically equivalent to the exemplified configuration, another configuration that includes a portion that is logically equivalent to the illustrated configuration, another configuration that is logically equivalent to the main part of the illustrated configuration, and the like are possible. is there. Further, another configuration that achieves the same or similar purpose as the illustrated configuration, another configuration that achieves the same or similar effect as the illustrated configuration, and the like are possible.
In addition, various variations of various components illustrated in the embodiment of the present invention can be implemented in appropriate combination.
Further, the embodiments of the present invention include inventions according to various viewpoints, stages, concepts, or categories, such as inventions of the entire device, inventions of components inside the device, or inventions of methods corresponding thereto. It is inherent.
Therefore, the present invention can be extracted from the contents disclosed in the embodiments of the present invention without being limited to the exemplified configuration.
[0074]
The present invention is not limited to the embodiment described above, and can be implemented with various modifications within the technical scope thereof.
[0075]
【The invention's effect】
According to the present invention, it is possible to optimize the neural network coupling weight based on a general model evaluation function, which is better than the conventional optimization.
[Brief description of the drawings]
FIG. 1 is a diagram showing a configuration example of a neural network connection weight optimization apparatus according to an embodiment of the present invention.
FIG. 2 is a flowchart showing an example of a processing procedure of the neural network connection weight optimization device according to the embodiment;
FIG. 3 is a view for explaining the processing procedure of the neural network connection weight optimization device according to the embodiment;
FIG. 4 is a diagram for explaining processing of a neural network element encoding unit of the neural network connection weight optimization apparatus according to the embodiment;
FIG. 5 is a diagram for explaining encoding function conversion in a conventional method;
FIG. 6 is a diagram for explaining coding function conversion in the neural network connection weight optimization device according to the embodiment;
FIG. 7 is a diagram for explaining inverse function conversion in the neural network connection weight optimization device according to the embodiment;
FIG. 8 is a diagram showing an example of the structure of a solution candidate neural network
FIG. 9 is a diagram showing an example of a model evaluation function serving as a standard for optimization
FIG. 10 is a diagram for explaining an example of optimization results in comparison with a conventional method;
FIG. 11 is a diagram for explaining an example of optimization results in comparison with a conventional method;
[Explanation of symbols]
101 ... Solution candidate neural network set storage unit
102 ... Model evaluation function storage unit
103: Neural network element encoding unit
104: Coding neural network set storage unit
105 ... New encoding neural network generation unit
106: New encoding neural network set storage unit
107: Neural network element decoding unit
108 ... New neural network set storage unit
109 ... New neural network evaluation / selection unit

Claims (16)

A neural network connection weight optimization method for optimizing a connection weight of a neural network having a predetermined connection structure composed of a plurality of network elements and zero or one or more bias elements,
One or more neural network, a first step of storing the connection weights of the neural network in the candidate solution neural network set storage unit in association with the evaluation results of the evaluation on the basis of a predetermined evaluation function,
For each of a predetermined number of neural networks selected from the neural network stored in the solution candidate neural network set storage means, for each connection weight related to input to the same network element among all the connection weights of the neural network, A second step of performing encoding by performing a predetermined encoding function conversion in consideration of the function of the network element and concatenating the codes obtained for each network element to generate an encoded neural network;
A third step of newly generating a predetermined number of encoded neural nets based on the encoded neural net generated in the second step;
For each of the encoded neural nets generated in the third step, a predetermined inverse considering the function of the network element for each code part corresponding to each network element among all the codes of the encoded neural network. A fourth step of performing function conversion to newly generate a neural network of the same format as the neural network stored in the solution candidate neural network set storage means ;
Each of the neural networks generated in the fourth step is evaluated based on the model evaluation function to obtain an evaluation result, and at least based on the evaluation result, the neural network and its evaluation result are determined as the solution candidate neural network. A fifth step of determining whether to add to the net set storage means and associating the neural network determined to be added with the evaluation result in association with each other and storing them in the solution candidate neural net set storage means A neural network connection weight optimization method characterized by: In the fifth step, the existing neural network stored in the solution candidate neural network set storage unit and the evaluation result satisfying a predetermined condition are deleted from the solution candidate neural network set storage unit. The neural network connection weight optimization method according to claim 1, wherein: In the fifth step, in the second step of the evaluation result of the neural network generated in the fourth step and the existing neural network stored in the solution candidate neural network set storage unit , The evaluation result of the selected neural network is compared, and the one having the better evaluation result is selected in the same number as the number of the neural network selected in the second step, and the selected neural network and The neural network stored in the solution candidate neural network set storage means and the portion of the neural network and the evaluation result selected in the second step among the evaluation results are updated with the evaluation result. The neural network connection weight optimization method according to claim 1. The second step to the fifth step are repeatedly executed until it is determined that a predetermined end condition is satisfied after the fifth step,
If it is determined that a predetermined termination condition is satisfied after the fifth step, among the neural networks stored in the solution candidate neural network set storage unit , the best evaluation result at that time The neural network connection weight optimization method according to any one of claims 1 to 3, wherein the neural network is selected as an optimized neural network. In the second step, when the predetermined encoding function conversion is performed on the connection weight related to the input to the same network element of the neural network, the total connection weight related to the input to the network element is calculated. The vector length of a coupling weight vector composed of coupling weights other than the coupling weights related to the input from the bias element is obtained, and all the vectors related to the input to the network element based on the vector length or a value obtained by multiplying the vector length by a positive number The encoded neural network is generated by concatenating a value obtained by performing a predetermined operation on each of the connection weights and the vector length or a value obtained by multiplying the vector length by a positive number as a code. The neural network connection weight optimization method according to any one of claims 1 to 4. 6. The neural network connection weight optimization method according to claim 5, wherein the predetermined operation is an operation of dividing the connection weight by the vector length or a value obtained by multiplying the vector length by a positive number. In the second step, when the predetermined encoding function conversion is performed on the connection weight related to the input to the same network element of the neural network, the total connection weight related to the input to the network element is calculated. Of these, the vector length of a coupling weight vector composed of coupling weights other than the coupling weight related to the input from the bias element is obtained, and the coupling weight related to the input from the bias element based on a value obtained by multiplying the vector length by a first positive number And a value obtained by performing a predetermined operation on each of the connection weights other than, and the vector length multiplied by a second positive multiple (however, the first positive number and the second positive number are different values and the first A value obtained by performing a predetermined operation on each of the coupling weights related to the input from the bias element based on a value obtained by at least one of a positive number of 1 and a second positive number being a value that is not 1), and the vector The length is the first positive 2. The encoded neural network is generated by concatenating, as a code, a set of a value obtained by specifying a value obtained by multiplying the multiplied value and a value obtained by multiplying the vector length by a second positive number as a code. 5. The neural network connection weight optimization method according to any one of items 4 to 4. The predetermined calculation divides a coupling weight other than a coupling weight related to an input from the bias element by a value obtained by multiplying the vector length by a first positive number, and a coupling weight related to an input from the bias element. 7. The method for optimizing a neural network connection weight according to claim 5, wherein the vector length is divided by a value obtained by multiplying the vector length by a second positive number. In the fourth step, in performing the predetermined inverse function transformation on a code portion corresponding to one network element of all codes of the coding neural network, the vector length or Based on a code value corresponding to a value obtained by multiplying the vector length by a positive number, a predetermined operation is performed for each code value corresponding to a value obtained by performing a predetermined operation on each of the coupling weights in the code portion. 7. The neural network connection weight optimization method according to claim 5, wherein the corresponding connection weight in the new neural network is obtained by applying The predetermined calculation obtains a correction value obtained by performing a predetermined correction on a code value corresponding to the vector length or a value obtained by multiplying the vector length by a positive number in the code portion, and the combination in the code portion. The neural network connection weight optimization method according to claim 9, wherein the correction is performed by multiplying a code value corresponding to a value obtained by performing a predetermined calculation on each of the weights. In the fourth step, when performing the predetermined inverse function transformation on the code portion corresponding to one of the network elements among all the codes of the coding neural network, the vector length of the code portion is set. The vector length specified based on a set of code values corresponding to a set of values that can specify a value obtained by multiplying a first positive number and a value obtained by multiplying the vector length by a second positive number is a first positive number. Based on a value corresponding to the multiplied value, a code value corresponding to a value obtained by performing a predetermined calculation on each of the coupling weights other than the coupling weight related to the input from the bias element in the code part. A code corresponding to a set of values that perform a predetermined calculation and that can specify a value obtained by multiplying the vector length by a first positive number and a value obtained by multiplying the vector length by a second positive number in the code portion. Identified based on value pairs Further, based on a value corresponding to a value obtained by multiplying the vector length by a second positive number, it corresponds to a value obtained by performing a predetermined calculation on each of the coupling weights related to the input from the bias element in the sign portion. 9. The neural network connection weight optimization method according to claim 7 or 8, wherein a corresponding connection weight in the new neural network is obtained by performing a predetermined operation on each of the code values. The predetermined calculation obtains a first correction value obtained by performing a predetermined correction on a value corresponding to a value obtained by multiplying the specified vector length by a first positive number, and the bias element of the sign portion is obtained. And multiplying the code value corresponding to the value obtained by performing a predetermined calculation on each of the connection weights other than the connection weights related to the input from the first correction value, and the specified vector length A second correction value obtained by performing a predetermined correction on a value corresponding to a value that is a multiple of 2 is obtained, and a predetermined calculation is performed on each of the coupling weights related to the input from the bias element in the sign portion. 12. The method for optimizing a neural network connection weight according to claim 11, wherein the calculation is performed by multiplying the code value corresponding to the value obtained by the second correction value. Said predetermined correction, the value that takes the absolute value correction, or the value is a positive number of threshold values determined the advance said value if it is less than the threshold value of a positive number that is predetermined for The neural network connection weight optimization method according to claim 10 or 12, characterized in that: The neural network connection weight optimization method according to claim 1, wherein in the third step, the encoded neural network is generated by a genetic algorithm. A neural network connection weight optimization device for optimizing a connection weight of a neural network having a predetermined connection structure comprising a plurality of network elements and zero or one or more bias elements,
One or more neural networks, and the solution candidate neural net set storage means for storing the connection weights of the neural network in association with the evaluation result of the evaluation on the basis of a predetermined evaluation function,
For each of a predetermined number of neural networks selected from the neural network stored in the solution candidate neural network set storage means, for each connection weight related to input to the same network element among all the connection weights of the neural network, Means for generating a coded neural network by performing a predetermined coding function conversion in consideration of the function of the network element and coding, concatenating the codes obtained for each network element;
Means for newly generating a predetermined number of encoded neural nets based on the encoded neural net generated by the means;
For each of the encoded neural nets generated by this means, a predetermined inverse function transformation taking into account the function of the network element is performed for each code part corresponding to each network element among all the codes of the encoded neural network. And a means for newly generating a neural network of the same format as the neural network stored in the solution candidate neural network set storage means ,
Each of the neural nets generated by the means is evaluated based on the model evaluation function to obtain an evaluation result, and at least the neural network and the evaluation result are stored in the solution candidate neural net set storage based on the evaluation result. Determining whether to add to the means, and means for associating the neural network determined to be added with the evaluation result and storing it in the solution candidate neural network set storage means, Neural network connection weight optimization device. A program for causing a computer to function as a neural network connection weight optimization device that optimizes connection weights of a neural network having a predetermined connection structure composed of a plurality of network elements and zero or one or more bias elements,
One or more neural network, a function for storing the connection weights of the neural network in the candidate solution neural network set storage unit in association with the evaluation results of the evaluation on the basis of a predetermined evaluation function,
For each of a predetermined number of neural networks selected from the neural network stored in the solution candidate neural network set storage means, for each connection weight related to input to the same network element among all the connection weights of the neural network, Encoding by performing a predetermined encoding function conversion in consideration of the function of the network element, concatenating the codes obtained for each network element, and a function for generating an encoded neural network,
Based on the encoded neural net generated by this function, a function for newly generating a predetermined number of encoded neural nets,
For each of the encoded neural nets generated by this function, a predetermined inverse function conversion considering the function of the network element is performed for each code part corresponding to each network element among all codes of the encoded neural network. And a function for newly generating a neural network of the same format as the neural network belonging to the above stored in the solution candidate neural network set storage means ,
Each of the neural nets generated by this function is evaluated based on the model evaluation function to obtain an evaluation result, and at least the neural network and the evaluation result are stored in the solution candidate neural net set storage based on the evaluation result. Determining whether to add to the means , and causing the computer to realize a function for associating the neural network determined to be added with the evaluation result and storing it in the solution candidate neural network set storage means program.

Images