JP3197377B2 - Neurocomputer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a neurocomputer,
In particular, the present invention relates to a neurocomputer which is applied to an image recognition device or the like using a neural network and has an optimized structure of the neural network.

[0002]

2. Description of the Related Art Hitherto, in an image recognition apparatus using a hierarchical neural network, the number of connections has been extremely large since the connections between neuron layers are often calculated as all connections in most cases. For example, FIG. 9 is a conceptual diagram showing a general connection configuration of a neural network of a conventional image recognition device. In the figure, input image 1
1 has been recognized through the neural network of the input layer 12, the intermediate layer 13, and the output layer 14, and the result has been output. As is clear from the figure, the input layer 12, the intermediate layer 1
3. The connection between the neuron layers of the neural network of the output layer 14 is fully connected.

The conventional image recognition apparatus in which the connection between the neuron layers is almost completely connected has a problem that it takes time to process the connection relationship. To address this problem, for example, John,
W. M. Robert. F. P. (1991) "Fractally
Neural Networks. Neural Networks, Vo
l.4.pp.53-60. Has been proposed to reduce the number of bonds. This is a method of determining the connection pattern of the neural network using fractals, for example, because the brain of an organism is structured to some extent as it was born. In other words, since a complex connection pattern can be described by a simple equation using fractals, there is a possibility that the connection pattern can be obtained at a very low cost.

A method of obtaining a connection pattern of a neural network by this method will be described. First, a series of x ₁ is sequentially obtained according to a simple equation as shown in equation (1).
The connection between neurons is determined according to the sequence. Here, equation (1) represents an equation for sequentially determining the order of inputs to neuron nodes in a certain layer. FIG. 10 is a conceptual diagram showing a connection configuration of a neural network of a conventional image recognition apparatus for determining a connection pattern of a neural network using this fractal. FIG. The following describes how to do this.

[0005] In the figure, an input image 15 corresponds to an input layer 1.
6, the intermediate layer 17, and the output layer 18 are recognized through the neural network, and the result is output. First,
When sequentially obtaining a series of x ₁ according to a simple equation as shown in equation (1), an appropriate initial value is _first given to x ₁ ⁽⁰⁾ . Thereafter, x ₁ ⁽¹⁾ and x ₁ ^(2
), And so on, the sequence of x1 is obtained one after another.

[0006] ^{_{x 1 (t + 1) ←}} c 1 (1- | x 1 (t) |) -1 ····· (1) This x1 sequence, for example, the initial value is set to 0.10, the c ₁ 1.5
Assuming that 0.10, 0.35, -0.03,..., Etc. are obtained in turn, this is normalized and replaced with the neuron number. For example, assuming that there are 100 neurons, the first connection is a connection with neuron number 10, the second connection is a connection with neuron number 35, and the third connection is a connection with neuron number 3. Thus, the connection order is determined in such a manner as to have a negative weight. Here, the first, second, and third lines are expressed for convenience, but are expressed in this way in consideration of the correspondence with the output sequence described below. If the sequence of input to a certain layer is 10,35, -3 …… and the sequence of output from the previous stage of that layer is supposed to be calculated as 88, -5,12, …… This means that the neuron number 88 of the previous layer
Is connected to the neuron number 10 of the subsequent layer, the neuron number 5 of the preceding stage and the neuron number 35 of the subsequent stage are connected with negative weight, and the neuron number 12 of the preceding stage and the neuron number 3 of the subsequent stage are also connected with negative weight. It indicates that the connection order is determined such as connection.

Next, when obtaining the order of outputs from a certain layer, it is necessary to obtain the same procedure as that for obtaining the order of inputs described above, except that equation (2) is used instead of equation (1). it can. Therefore, the description of the method is omitted here.

X ₂ ^{(t + 1)} ← c ₂ (1− | x ₂ ^(t) |) −1 (2) By introducing a connection rule between neuron layers by repeatedly using this simple equation In addition, a fractal structure can be constructed in the connection relation of each layer of the neural network.
However, it is usually very difficult to appropriately determine the parameters (c ₁ , c ₂ ) determined for each layer in the above-described method. These parameters (c ₁ , c ₂ ) can be determined in reverse by the above-described sequence for determining neurons between the preceding and succeeding layers. However, conventionally, the parameters (c ₁ , c ₂ ) are determined by optimizing the parameters from the sequence describing this connection relationship. In this case, a simulated annealing method was used. However, as the number of layers increases or the number of inputs increases, the degree of freedom of the system increases, and the connection relationship between neurons enters a large degree of arbitrariness. there were. On the other hand, although attempts have been made to determine parameters using a genetic algorithm for the annealing part, the amount of calculation has similarly become a bottleneck.

FIG. 11 is a conceptual diagram showing a connection configuration of a neural network of a conventional image recognition apparatus for determining a connection relationship between neurons by using a genetic algorithm for this annealing. The method of determining the connection relationship between neurons using the genetic algorithm of FIG. 11 will be described together with the description of FIGS. 12 and 13 described later. However, in FIG. 11, similarly to FIG. 10, the input image 15 is recognized via the neural network of the input layer 16, the intermediate layer 17, and the output layer 18,
The result is output.

Since the conventional neural network of the image recognition apparatus is such a system having a large amount of calculation, a conventional example which refers to a method of extending to a multilayer in which the amount of calculation is further increased to improve the recognition rate is described. Did not.

Next, FIG. 12 is a conceptual diagram for explaining an image recognition operation of an image recognition device using a neural network which determines a connection relationship between neurons using a genetic algorithm. In the figure, a conventional image recognition device using a neural network inputs an input image 21 and an input layer 22 in which neurons are arranged on a plane.
And a middle layer 23 for extracting features of image data, and an output neuron group 24 for separating various features and finally recognizing what the input image is. , The final recognition result of the input image is output from the output neuron group 24. For the sake of simplicity, the recognition process will be described below on the assumption that the input image is an image of an airplane having various shapes.

First, the description will be made by limiting the number of intermediate layers to one. In FIG. 12, the intermediate layer 23 represented by a parallelogram
On each plane ₂₃ 1 ~ 23 _m that are configured neurons around. Hereinafter, this density of neurons will be referred to as a feature plane. Wherein plane 23 ₁ ~ 23 _m is a neuron population to extract each partial image features of the input image. For example, in FIG. 12, wherein plane 23 ₁
Is intended to detect the shape of the wing sections in the plane image, wherein the plane 23 ₂ is intended to detect the shape of the tail portion of the image, wherein the plane 23 ₃ detects the shape of the nose portion of the image Is what you do. Similarly, none of the neuron population in the feature plane, wherein the plane 23 ₁ ~ 23 _m, the different parts of the input plane image, it is assumed that the population of extracting any of image features.

[0013] To these characteristics planes 23 _1-23 convenience symbol a is a _m below, b, c, d, ......... denoted. Assuming that the connection state of the neurons on the feature plane can be represented by ternary values, the feature plane 23 ₁ (symbol a) becomes “0,
1,0,1, -1 ... ”. However, here, for convenience of explanation, the symbol a (characteristic plane 23
It is assumed that the contents of the feature plane are collectively represented by the notation of 1). Now, feature plane 23 ₁
〜23 _m are all connected in a row as [abcd...] To generate a gene. Then, as shown in FIG. 13, which will be described later, this gene generation processing is also performed for other neural networks, and gene crossover is performed using genes generated from intermediate layers at the same position in a plurality of neural network systems.

In this case, how similar the contents of each feature plane are can be obtained by measuring the similarity of the symbols a, b, c, d,.... For example, a and a '
Are similar and a and d are hardly similar.
In other words, the same symbol to which "'" (dash) is added indicates that the connection state in each feature plane is similar, and that the feature planes of different symbols are not completely similar.

In the case of gene crossover using a conventional genetic algorithm, generally, even if there is a feature very similar to that of the parents, the similar feature is not always inherited by the child. This genetic mechanism will be described with reference to FIG. FIG. 13 is a conceptual diagram showing a conventional gene crossover procedure, and shows a method of generating a child gene by crossing parents' genes.

In FIG. 13A, the gene A of the individual A, which is the parent gene before crossover, obtained from the intermediate layers of the separate neural networks, and the gene 26 of the individual B
It is shown. As shown in the figure, the contents of the feature plane of the gene 25 of the individual A are “abcdef”, and the contents of the feature plane of the gene 26 of the individual B are “d ′”.
gb "a'f" c "". FIG. 13B shows a method in which the parent genes 25 and 26 are simply crossed over according to a conventional gene crossing method. FIG.
(C) shows the gene of the child obtained as a result of the gene crossover. As shown in the figure, the content of the feature plane of the gene 27 of the individual C, which is the child gene generated by the crossover, is “abca'f” c ””, and the content of the feature plane of the gene 28 of the individual D is , "D'gb" def "
It is. When the parent gene is simply crossed according to the conventional gene crossover method, the gene of the child neural network finally obtained is, for example, as shown in FIG.
Similar patterns such as “a” and “a ′” are transmitted redundantly like the gene 27 of the individual C shown in (c), or a part of the feature plane of the parent's gene is like the gene 28 of the individual D. I can't tell at all.

The fact that there is a very similar part in the genes of the parents means that during the process of extracting features by repeating mating several times, the features are generally useful to the system to some extent. Means that. Therefore, it is not preferable in terms of performance that the child system inherits this universal feature redundantly, or conversely, does not inherit it at all.

[0018]

As described above, in the conventional neural network system, methods for optimizing the structure of the connection relation of neurons include a method using a simulated annealing method and a genetic algorithm. However, in each case, there is a problem that the amount of calculation of the connection relationship becomes enormous and a long time is required for processing. In order to improve the recognition rate, it is conceivable to extend the neural network in multiple layers.However, the method using the genetic algorithm is a conventional example that mentions such a processing method because the amount of calculation of the connection relation is huge. There was no. Furthermore, in the gene crossover process of the conventional genetic algorithm, since the genes of the parents were simply crossed over, similar patterns are duplicated and conveyed in the finally obtained gene of the child neural network. Alternatively, there was a problem that some feature planes of the genes of the parents were not transmitted at all.

The present invention has been made in order to solve the above-mentioned problems, and realizes a genetic algorithm that enables generation of a child gene that optimally inherits the characteristics of the parents' genes, and also reduces the amount of calculation. We realized a genetic algorithm that can suppress the increase as much as possible, and optimized the connection relation of the neural network that can efficiently determine the connection pattern between each layer including the case where the intermediate layer of the neural network is expanded to multiple layers. The aim is to get a neurocomputer.

[0020]

In order to achieve the above object, a neural computer according to a first aspect of the present invention uses a plurality of hierarchical neural networks independently of each other, and uses an intermediate layer of each neural network. Gene generating means encoding a connection rule of a neural network for each group of neurons for extracting the features in a gene generating means, and a gene generating means using the constituent element sequence connecting them as a gene, and gene generating means Sorting means for classifying and sorting the components included in the genes according to their functions, and selecting two genes from the number of neural networks generated by the gene generating means,
Genetic crossover means for sorting and comparing the selected constituent elements of both genes by rearranging means to crossover so that a child gene that optimally inherits the characteristics of the genes of the parents can be generated. The connection rule between the intermediate layer and the input / output layer before and after the intermediate layer is optimized by a genetic algorithm by the means.

Further, neuro-computer according to the second invention, in the neuro-computer using a two-menu neural network, using a hierarchical neural network into a plurality each independently, characterized in the intermediate layer of each neural network Gene generation means for generating a gene component encoding a connection rule of a neural network for each group of neurons for extracting a gene, and a gene sequence comprising the component sequence connecting them, and a gene generated by the gene generation means are included. Sorting means for sorting the components to be sorted in accordance with the similarity , and sorting by the sorting means.
Crossed genes to produce child genes
And genetic crossover means formed, for each intermediate layer, at the layer
A conversion table that stores information on the rearrangement of gene components by gene crossover means , and gene crossover in a certain intermediate layer
When performing, the conversion table for the relevant layer and the middle of the previous stage
Performs re reading genetic components between the conversion table for the layer, characterized in that a means to recover the destruction of recombinant accompanied the Hare binding relationship of the components by genetic crossover in the layer.

That is, the neural computer according to the present invention uses a plurality of hierarchical neural networks having a plurality of intermediate layers independently of each other. (1) Each intermediate layer of each neural network sequentially (2) Each intermediate layer of each neural network is characterized in that it is structured so as to sequentially extract higher-order integrated features from the foremost stage to the last layer. In order to optimize the connection relationship with, a gene component encoding the connection rule of the neural network is generated for each group of neurons for extracting features in the hidden layer,
A gene generating means having a sequence of constituent elements connecting them as a gene, a rearranging means for classifying and sorting the components included in the gene generated by the gene generating means according to the similarity , and a sorting means which is sorted by the rearranging means.
Genetic crossover means for crossing genes to generate a child gene realizes an optimized genetic algorithm. (3) Each intermediate layer of each neural network has a gene crossover in that layer. performs a conversion table storing recombinant information components, the re-read of the components between the conversion table of the conversion table and the previous stage of the layer of the layer when performing gene crossover in the layer in, in the layer Means for restoring the destruction of the binding relationship up to that time due to the rearrangement of the constituent elements due to the gene crossover.

[0023]

Therefore, according to the neurocomputer of the present invention, gene components encoding the connection rules of the neural network are generated for each group of neurons for extracting features in the intermediate layer by the gene generation means, and are connected. The components included in the gene can be classified and sorted by function by the rearrangement means, and a child can be generated that optimally inherits the characteristics of the genes of the parents selected from the genes of the number of neural networks by the gene crossover means. In this way, the possibility of generating a neural network with clearly poor performance from the beginning can be reduced from the beginning, making it unnecessary to calculate unnecessary connections between neurons. Can be

Further, according to the neurocomputer of the present invention, the above-mentioned optimal layer is obtained by each intermediate layer of each neural network structured so as to successively extract high-order integrated features from the first stage to the last layer. it is possible to realize a reduction by genetic algorithms, each intermediate layer stores recombinant information elements in genetic crossover in the layer by the conversion table, the conversion table and the previous stage of the layer of the layer when performing genetic crossover Since the components are exchanged with the conversion table, it is possible to restore the previous destruction of the connection relationship due to the rearrangement of the components due to gene crossover in the relevant layer, and to speed up the optimization of the neural network. It can be carried out. Therefore, even when it is necessary to increase the number of identification data and improve the recognition rate by increasing the number of layers of the neural network, an optimized genetic algorithm is realized to optimize the connection relationship between the layers of the neural network. Can be easily performed.

[0025]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below with reference to the drawings.

Embodiment 1 FIG. FIG. 1 is a conceptual diagram showing a flow of a process of finally selecting one neural network by using the genetic algorithm according to the present embodiment in optimizing the connection between hidden layers of the neural network in the neural computer according to the present embodiment. FIG.

First, a general genetic algorithm will be described. Genetic algorithms, as the name implies,
It introduces a survival tactic in the world of living things, in particular, a mechanism in which living things have the optimal mechanism for the external environment while repeating evolution and selection, and survive as a selection principle of general systems other than living things. In other words, when there are several systems that perform a similar function, each system is analyzed first, and what is characteristic of each system is examined (here, each system described above is referred to as a genetic algorithm). Then it is called "individual"). For example, to give a simple example, suppose there is an automobile manufacturer that wants to sell a new car. In this case, the features of the system correspond to the fact that the new car has a new engine, is well-designed, or has a new braking mechanism. And the rating of this new car is determined by how many people buy it, as is the law of natural selection. Now, assuming that a new car named A sold explosively, many of the features of Car A would be taken over in the development of the next generation of new products. Conversely, car A
If the car does not sell well, the characteristics of the car A will hardly be carried over to the development of a new product of the next generation. This is realized by the genetic algorithm as follows.

That is, if the features of the car A are listed as a gene a and the features of another car B being sold at the same time are set as a gene b, the genes are arbitrarily mixed to generate the next generation. Develop cars C, D, E, ... Cars C, D, E,... Only those cars actually sold in the market by a certain number or more are qualified to be parents of the next generation. By repeating this genetic rule for generations, it is possible to create a car with good properties that inherits genes that sell very well in the world. The above is the outline of the general genetic algorithm.

However, as described in the prior art, when the parent gene is simply crossed over according to the conventional gene crossover method, the finally obtained gene of the child neural network is obtained by transmitting similar patterns of the parents in duplicate. There was a problem that some feature planes of the parents' genes were not transmitted at all. The genetic algorithm of the neurocomputer of the present embodiment aims at solving such a problem.

First, FIG. 1 is a conceptual diagram showing a processing flow of the neurocomputer of this embodiment for finally selecting one neural network by using the genetic algorithm for optimizing the connection of the hidden layers of the neural network. is there. The figure shows that neural networks having different connections of the intermediate layer one by one are formed for each generation, and a neural network having an excellent result is gradually selected from the neural networks. The final selection is the one that has performed well for a given process, shown as the last generation in the figure.

Next, the procedure and means of gene crossover for realizing the genetic algorithm of the present embodiment in which the neural network with the best performance is selected will be described with reference to the drawings. First, FIG. 2 is a conceptual diagram showing a procedure of gene crossover in the neurocomputer of the present embodiment, and shows a method of crossing genes by sorting them. The difference from the conventional gene crossover procedure shown in FIG. 13 described in the prior art is that the genes of the parents are almost equally inherited by sorting and sorting the genes of the parents by their similarity before performing the gene crossover. That is to make the gene of the child. This will be specifically described below.

First, FIG. 2 (a) shows the gene 35 of the individual A and the gene 36 of the individual B, which are the parent genes before crossing, obtained from the intermediate layers of different neural networks, respectively. . The content (component) of the characteristic plane of the gene 35 of the individual A is “abcdef”, and the content of the characteristic plane of the gene 36 of the individual B is “d′ gb”.
a'f "c"'. FIG. 2B shows a process of rearranging the components of the parent genes 35 and 36 in descending order of similarity. FIG. 2C shows a method in which parent genes 35 and 36 rearranged in descending order of the similarity are crossed in accordance with a conventional crossover method. Further, in FIG.
Genes of children obtained as a result of gene crossover are shown. As shown in FIG. 2D, the content of the characteristic plane of the gene 37 of the individual C, which is the child gene generated by the crossover, is “adbf” c ”g”, and the characteristic plane of the gene 38 of the individual D. Is "a'd'b" fce "
It is. In FIG. 2B, sorting is performed in ascending order of similarity, but the similarity is calculated by obtaining the Euclidean distance between the constituent elements of the genes of the parents.

In addition, gene crossover can generally be performed at a site of an arbitrary length, whereby it is possible to generate genes of a plurality of children in which the mixing ratio of the genes of the parents is changed. Therefore, it was explained that gene crossover is performed at the center of each of the parents' genes.

When the genes of the parents are sorted and crossed as described above, the genes of the child neural network finally obtained are, for example, a and a like the gene 27 of the individual C shown in FIG. Similar patterns such as a 'are not transmitted in duplicate, and processing calculations relating to the connection rules of the neural network accompanying the duplication can be omitted. Further, since some feature planes of the genes of the parents are not transmitted at all like the gene 28 of the individual D shown in FIG. 13C, there is a lack of some important feature planes. Processing calculations relating to the connection rules of the neural network can be omitted, and the structure of the neural network can be optimized at high speed.

Next, means for implementing the genetic algorithm of FIG. 2 will be described. The neural computer according to the present embodiment is obtained by coding a neural network connection rule for each group of neurons for extracting features in an intermediate layer of each neural network by using a plurality of hierarchical neural networks independently. Gene generating means for generating (components of a gene), and using a sequence of constituent elements connecting them, and sorting means for classifying and sorting the components included in the gene generated by the gene generating means according to their functions. And two genes are selected from the number of neural networks generated by the gene generating means, and the components of both selected genes are compared, and a pair of components showing the same output with respect to the same input is selected. Gene crossover so that these are transmitted to different children, and optimally receive the characteristics of the parents' genes. It is and a gene crossover means that succeeded's children gene has to be able to generate.

Hereinafter, the flow of the procedure of the gene crossover process of the neurocomputer of this embodiment will be further described with reference to the flowchart shown in FIG. In the figure, in order to optimize the connection rules between the intermediate layer of each neural network and the layers before and after it by a genetic algorithm, first, a neural network A and a neural network B as parents are selected from a set of neural networks. Select (step 51). Next, a gene component encoding a feature plane in the intermediate layer of the neural network A and the neural network B selected by the gene generating means is generated, and a component string connecting these is defined as a gene (step 52). . The Euclidean distance between the constituent elements is determined based on the genes obtained from both neural networks (step 5).
3). The rearranging means rearranges the constituent elements in ascending order of similarity based on the obtained Euclidean distance between the constituent elements (step 54), and the gene crossover means crosses the sorted genes (step 5).
5). The neural network C, which is a child individual who has inherited the characteristics of both parents appropriately by this gene crossover,
D is generated (step 56). In the gene crossover process, the gene crossover means compares the genes of the parents that are the basis of the crossover, detects a pair of gene components in the feature extraction layer that shows the same output for the same input, and Is transmitted to different children.

Embodiment 2 FIG. Embodiment 1 described above. In the above, we explained the case where the number of hidden layers in the neural network is 1.
Next, a case where the number of intermediate layers is plural will be described. When the intermediate layer of the neural network is extended to multiple layers, if the genetic algorithm of the first embodiment described above is applied as is to separate neural networks (individuals), the connection will extend between individuals, so Then, the bonding relationship before and after each layer collapses. In the neurocomputer according to the second embodiment, the information on the rearrangement of the components in the gene crossover of each layer is held using a conversion table as shown in FIGS. The components are read between the conversion table of this layer and the conversion table of the relevant layer, and the connection between the preceding and succeeding layers is restored by means for restoring the disruption of the conventional connection relationship due to the rearrangement of the components due to the gene crossover in the relevant layer. I am able to maintain relationships.

FIG. 4 is a conceptual diagram for explaining an image recognition operation of an image recognition apparatus using a neural network having a plurality of intermediate layers for determining a connection relationship between neurons using the genetic algorithm of this embodiment. is there. In the figure, an i-th intermediate layer (i-th layer) of a plurality of intermediate layers of a certain neural network performs gene crossover with the i-th intermediate layer of another neural network. As a result, the rearrangement of the component rows generated in the i-th intermediate layer is of course the i + 1-th intermediate layer (the i + 1-th intermediate layer).
Layer) must be taken into account when rearranging. The maintenance and restoration of this connection relationship is realized using a conversion table.

FIGS. 5 to 8 are conceptual diagrams for explaining a processing procedure for restoring the conventional destruction of the connection relationship due to the rearrangement of components due to gene crossover using the above-described gene rearrangement information. First, in FIG. 5, it is assumed that the neural network A and the neural network B have the connection relationship between the i-th layer and the (i + 1) -th layer already constructed as shown in the figure. In this state, if the neural network A and the neural network B cross at the i-th layer, the connection relationship between the i-th layer and the (i + 1) -th layer which has been connected up to that time is straddled by the neural networks A and B. Is broken.
Therefore, in the neurocomputer of this embodiment, the state after the gene crossing of the i-th layer is stored in the conversion table 44 shown in FIG. 6, and the state is reflected and combined when performing the gene crossing in the next (i + 1) -th layer. They try to maintain and restore relationships. In the conversion table 44 in FIG. 6, genes replaced by gene crossover in the i-th layer are distinguished from those that are not by 1/0 identifiers. That is, the identifier 1 indicates that gene replacement is performed by gene crossover in the relevant layer,
Identifier 0 indicates that no gene replacement is performed by gene crossover in the layer.

FIG. 7 shows a conversion table 45 based on gene crossover in the (i + 1) th layer. here,
In each column of the conversion table 45, "g (a, b), u (o,
t)... ”means that the components g and u are the components a and b of the preceding layer (the i-th layer) and the components o and t, respectively.
Is connected to Hereinafter, component u
The process of maintaining and restoring the connection relationship by gene crossover will be specifically described with reference to FIG. Conversion table 45 shown in FIG.
Therefore, the component u does not change between the neural networks A and B in the gene crossover of the (i + 1) th layer,
The identifier is 0. Next, according to the conversion table 44 shown in FIG. 6, the components o and t of the previous layer connected to the component u have not been replaced with the component o (that is, the identifier is 0). Conversely, it can be seen that the component t is replaced (that is, the identifier is 1). Therefore, in the gene crossover of the (i + 1) th layer, it is necessary to replace the component t, and the component t is replaced with the component f from the conversion table 44.

By repeating the above procedure, even if gene crossover occurs, gene crossover can be easily performed without breaking the connection relationship of the neural network.

[0042]

As described above, according to the neurocomputer of the present invention, the gene component encoding the connection rule of the neural network for each group of neurons from which the features in the hidden layer are extracted by the gene generating means is obtained. The components that are generated and included by the gene that connects them are
Sorting and sorting for each function by sorting means,
The gene crossover means is designed so that gene crossover can be performed, which allows the generation of children that optimally inherits the characteristics of the parents' genes selected from the number of genes in the number of neural networks. The possibility of generation can be reduced as compared with the related art, and there is no need to calculate unnecessary connection relationships between neurons, so that there is an effect that optimization can be performed at high speed. Therefore, there is a high possibility that a higher recognition rate can be achieved within a certain period of time.

Further, according to the neurocomputer of the present invention, the above-mentioned optimal layer is obtained by each intermediate layer of each neural network structured so as to sequentially extract a high-order integrated feature from the foremost stage to the last layer. it is possible to realize a reduction by genetic algorithms, each intermediate layer stores recombinant information elements in genetic crossover in the layer by the conversion table, the conversion table and the previous stage of the layer of the layer when performing genetic crossover Because the components are read between the conversion table and the disruption of the binding relationship up to that point due to the rearrangement of the components due to gene crossover in the relevant layer, it is configured to be an optimized multilayer neural network. Can be easily realized, and the optimization processing of the neural network can be performed at high speed. Therefore, even when it is necessary to increase the number of identification data and improve the recognition rate by increasing the number of layers of the neural network, the neural computer of the present invention makes it easy to optimize the connection relationship between the layers of the neural network. There is an effect that can be realized.

[Brief description of the drawings]

FIG. 1 shows that the genetic algorithm of the present embodiment is used for optimizing the connection of hidden layers of a neural network.
It is a conceptual diagram which shows the flow of a process which selects one neural network.

FIG. 2 is a conceptual diagram showing a procedure of gene crossover in a neurocomputer of the present embodiment.

FIG. 3 is a flowchart illustrating a flow of a gene crossover processing procedure of the neurocomputer according to the present embodiment.

FIG. 4 is a conceptual diagram for describing an image recognition operation of an image recognition device using a neural network having a plurality of intermediate layers for determining a connection relationship between neurons using a genetic algorithm.

FIG. 5 is a diagram for explaining a processing procedure for restoring the destruction of the connection relationship using the gene recombination information.

FIG. 6 is a diagram for explaining a processing procedure for restoring the destruction of the connection relationship using the gene recombination information.

FIG. 7 is a diagram for explaining a processing procedure for restoring the destruction of the connection relationship using the gene recombination information.

FIG. 8 is a diagram for explaining a processing procedure for restoring the destruction of the connection relationship using the gene recombination information.

FIG. 9 is a conceptual diagram showing a general connection configuration of a neural network of a conventional image recognition device.

FIG. 10 is a conceptual diagram showing a connection configuration of a neural network of a conventional image recognition device that determines a connection pattern of the neural network using a fractal.

FIG. 11 is a conceptual diagram showing a connection configuration of a neural network of a conventional image recognition device that determines a connection relationship between neurons using a genetic algorithm in an annealing part.

FIG. 12 is a conceptual diagram for describing an image recognition operation of an image recognition device using a neural network having one intermediate layer that determines a connection relationship between neurons using a genetic algorithm.

FIG. 13 is a conceptual diagram showing a conventional gene crossover procedure.

[Explanation of symbols]

 11, 15, 21 Input image 12, 16, 22, 41 Input layer 13, 17, 23, 42 Intermediate layer 14, 18, 24, 43 Output layer 25, 35 Gene of individual A (parent) 26, 36 Individual B ( Gene of parent 27,37 Gene of individual C (child) 28,38 Gene of individual D (child) 31 First generation neural network 32 Second generation neural network 33 Final generation neural network 44 Conversion of i-th layer Table 45 Conversion table of the (i + 1) th layer

Claims (2)

(57) [Claims] 1. A neural computer using a neural network, wherein a plurality of hierarchical neural networks are independently used, and a neural network connection rule for each group of neurons for extracting a feature in an intermediate layer of each neural network. Gene generating means for generating gene components encoding the above, and generating a sequence of component elements connecting them, and sorting means for sorting the gene components included in the gene generated by the gene generating device according to the degree of similarity And a gene crossing means for generating a child gene by crossing the genes sorted by the rearranging means, wherein the connection rules between the intermediate layer of the hierarchical neural network and the input / output layers before and after the intermediate layer are determined by the means. Optimized by genetic algorithm Neuro-computer according to claim. 2. A neural computer using a neural network, wherein a plurality of hierarchical neural networks having a plurality of intermediate layers are used independently of each other, and a group of neurons for extracting features in each intermediate layer of each neural network. Gene generation means that generates gene elements encoding the connection rules of the neural network for each layer, and connects them to gene elements for each intermediate layer, and genes generated by the gene generation means are included Rearrangement means for sorting gene components to be performed according to similarity, gene crossover means for generating child genes by crossing over genes sorted by the rearrangement means, gene crossover in each intermediate layer for each intermediate layer Memorizing information of rearrangement of gene components by means When performing a gene crossover in a certain intermediate layer, the gene component is replaced between the conversion table for the layer and the conversion table for the preceding intermediate layer, and the gene crossover in the layer is performed. Means for restoring the destruction of the connection relationship due to the rearrangement of the constituent elements.