JP6998561B2 - A method for constructing a machine learning model for ultra-deep regression analysis, its device, its program, and a general-purpose mobile terminal device equipped with the program. - Google Patents

Description

The present invention belongs to the field of artificial intelligence, and in particular, a method for constructing a machine learning model for ultra-deep regression analysis, a device thereof, a program thereof, and a general-purpose mobile terminal device equipped with the program.

Developed by Google's investment, AlphaGo's achievements in winning the world championship of the swordsman once again heightened the global enthusiasm for deep learning. The number of artificial intelligence patent applications over the past year has far exceeded the total number of applications over the past year.

However, the conventional neural network algorithm using deep learning needs to combine all the possibilities in order to obtain the optimum solution when selecting the weight value W and the threshold value T of the conventional neural network. The total number of combinations is P × (W × T) ⁿ , where P is the number of layers of the conventional neural network, n is the number of contacts per layer, and W or T is the number of parameters. And.

In addition, it was found that it is almost impossible to obtain the calculation result using the current computer because the calculation complexity becomes O (n ⁿ ) due to the huge number of combinations and it becomes an NP-complete problem. rice field.
Since the definition method of the weight value W and the threshold value T of the conventional neural network model uses only the conventional mathematics, as a result, it is completely different from the principle of the trigger signal of the brain. The nerves of the human brain have revealed a completely different mechanism from the conventional neural network model.

Moreover, since the actual objective function originally belonged to the problem of probability, the conventional neural network model can only map the function. Therefore, it is necessary to rely on mass learning data. It is difficult to solve the probability problem. In particular, when the weight value W and the threshold value T are selected, the optimum solution is not obtained, and even if the stochastic gradient descent SGD for the loss function is used, only the local optimum solution is obtained, so that the black box problem remains.

The deep learning model that is currently attracting attention has not actually changed from the conventional neural network model. All you have to do is increase the number of hidden layers by hundreds or more.
It is doubtful that such a method can really increase the processing power. It cannot be theoretically proved that there is a relationship between the number of hidden layers and the processing capacity.
Due to the above causes, the result seems to be more complicated in calculation.
As described above, it cannot be expected to be widely applied industrially as a deep learning model.

Therefore, Furukawa Denki Kogyo Co., Ltd. of Japan, which is well known, applied for a patent called "image processing method and image processing device" that selects the image processing threshold value using a neural network algorithm and extracts the outline of the image. Patent Document 1) has been filed.

For applications in the field of autonomous driving, Toyota, a major Japanese company, has submitted an application called "Driving Orientation Estimator" (Patent Document 2). This application uses the machine learning algorithm of the reverse transfer neural network to avoid the occurrence of traffic accidents and automatically to prevent unforeseen circumstances even if the driver is not reflected. We are proposing a vehicle control method that allows you to select.

In the field of application to image analysis, Japan's Hosei University has published a patent application (Patent Document 3) entitled "Plant Disease Diagnosis System, Plant Disease Diagnosis Method, and Program". It is a method of diagnosing a plant lesion by recognizing a leaf image of a plant by introducing.

Japan's Fuji Xerox, the world's largest copier, has filed a patent application for "counterfeit prevention devices and methods using micro security marks" from the perspective of anti-counterfeit identification. Introduced and with a small security mark, it is intended to achieve the purpose of resolving the identification of product security.

Japanese Unexamined Patent Publication No. 2013-109762 Japanese Unexamined Patent Publication No. 2008-225923 Japanese Unexamined Patent Publication No. 2016-168046 Japanese Unexamined Patent Publication No. 2012-124957

Since deep learning is used in [Patent Document 1], [Patent Document 2], and [Patent Document 3], it is difficult to improve problems such as Black Books due to the above.

Further, in the above (Patent Document 4), it is unclear what kind of artificial intelligence or means is used in the application for generating the anti-counterfeit mark by artificial intelligence, and what kind of result is achieved. be. Further, as a method of describing microdot information for a microsecurity mark, it has become a well-known technique to describe information by different dot sizes, different dot colors, and different dot positions. Yes, even if artificial intelligence is introduced here, it is not novel because it is the same as the conventional technology.
Also, in this method, the anchor mark for indicating the position or direction of the data mark uses a larger or different form than the data mark that can describe information, but the result of embedding the information is in a secret form. It is difficult to become.

In particular, it is inevitable that an illegal person will use a scanner to copy minute anti-counterfeit marks.Since the printing accuracy is lower than the accuracy of the scanner, even the smallest anti-counterfeit marks will be completely copied using the scanner. It is possible. This application also stated that a laser marker was used to create a minute anti-counterfeit mark, but even with such a method, it is inevitable that it belongs to the conventional known technology, and actually various methods are disclosed. It is desirable to have a method that can actually identify counterfeit products using a mobile phone, but this application does not provide such a method.
Hereinafter, in this application, the contents of the following definitions that define related technical terms belong to the scope of the invention.

Probability Space:
Based on the Soviet mathematician Andrey Kolmogorov's theory that "probability theory is based on measure theory", the so-called probability space is a measurable space (S, M) with a probability measure μ (S) = 1. It is called the entered measure space (S, M, μ).

Probability Distribution:
It describes the likelihood of each value of a random variable.

Probability Scale:
In any probability distribution in probability space, one probability scale must exist and is a measure of the degree of probability distribution.
Probability Density:
Integral value of probability distribution function in a given domain

離散式の公式：

Fuzzy Event Probability Measure (Probability Measure of Fuzzy Event)
In the Euclidean space S including the probability space, if μ _A (x) is a membership function and P (x) satisfies the property of a probability measure such as additiveity, then μ _A (x) is also additive. When the fuzzy measure such as is satisfied, the measure P (A) of the set A of the fuzzy event probability measures is

Discrete formula:

Intelligent System:
It is to build a system based on a deterministic algorithm. Also, it is to achieve a certain objective function based on one algorithm, and the result of the processing is a deterministic system.

Artificial Intelligence:
Artificial intelligence is the achievement of the function of the human brain simply by using a computer. In other words, it is trying to achieve the effect of human brain processing via a computer. That is, it is a problem of a complex system that solves an uncertain problem or a problem that cannot be predicted in advance.

Artificial intelligence is to model and formulate the effects of human intervention, in other words, artificial intelligence is to solve stochastic problems using machine learning models.

Clustering:
To obtain a data set non-autonomously based on the Euclidean spatial scale within a range given in advance.

Self-organization:
An algorithm that can autonomously transfer data in the direction of high probability based on the scale of probability space, or iteratively based on the model of formulation to obtain results that exceed the model of formulation. be.

Machine Learning:
A model in which a computer can autonomously acquire rules from prior data.

確率空間における、確率分布ｇ_ｆに対し、特徴関数Ａ（Ｇ）の値が存在すべきである。したがって、確率空間は測定空間であるため、特徴関数Ａ（Ｇ）に対して、必ず確率尺度関数Ｍ［Ｇ，Ａ（Ｇ）］が存在する。確率尺度を基準として、集合Ｇ^（ｎ）を最大確率方向に転移できる条件は数４のようになる。

ｎがρ（ρを４以上の値とする）より大きい場合になると、自己組織の結果として、目的関数に対して、最大確率分布値をＡ（Ｇ^（ｎ））とし、また、Ａ（Ｇ^（ｎ））を中心とした最大確率尺度をＭ［Ｇ^（ｎ），Ａ（Ｇ^（ｎ）］とすることができる。Self-organization based on a Probability Scale: Self-organization based on a probability scale:
Let G be a set of probability distributions in a probability space, and g _f be the data belonging to the set:

There should be a value of the characteristic function A (G) for the probability distribution g _f in the probability space. Therefore, since the probability space is a measurement space, the probability scale function M [G, A (G)] always exists for the characteristic function A (G). The condition that the set G ⁽ⁿ⁾ can be transferred in the maximum probability direction based on the probability scale is as shown in equation 4.

When n is larger than ρ (where ρ is a value of 4 or more), as a result of self-organization, the maximum probability distribution value is A (G ⁽ⁿ⁾ ) with respect to the objective function, and A (G). The maximum probability scale centered on ⁽ⁿ⁾ ) can be M [G ⁽ⁿ⁾ , A (G ⁽ⁿ⁾ ]].

Maximum Probability:
It is the probability value of the prediction that exceeds the result of the conventional statistical prediction and is the closest to the mother.

Probability Space Distance:
In the probability space included in the Euclidean space, v _j ∈ V (j = 1, 2, ..., N) is an element of the probability distribution, the maximum probability scale of the probability distribution is M _j , and r _j ∈ R is the Euclidean space. As an element of the set R in, the distance G (R, V) from the set R in the Euclidean space to the set V in the probability space can be defined as follows.

Super Deep Learning:
A new neural network model consisting of a sensing layer, a nerve layer, and a cerebral skin layer was constructed. Probability information is transmitted, where the points of contact between the input information and the sensing layer, the sensing layer and the nerve layer, and the dermal layer of the nerve layer are connected by unsupervised machine learning of self-organization based on a probability scale. Is. It is also possible to simulate the function of the brain.

A first object of the present invention is ultra-deep competitive learning, which makes it possible to obtain an optimum similarity relationship between data by ultra-deep competitive learning using minute unstable probability information and minute ambiguous information. Build a neural network model of. To provide the highest level of machine learning models for image recognition, voice recognition, etc., or industrial automation.

A second object of the present invention is to provide a method for finding a unified exact distance between the Euclidean space and the probability space.

A third object of the present invention is to provide a method for obtaining a strict fuzzy event probability measure in which the Euclidean space and the probability space are unified.

A fourth object of the present invention is to provide an ultra-deep regression analysis learning model capable of a more accurate functional approach than the conventional linear regression analysis method.

A fifth object of the present invention is to provide a model of ultra-deep competitive learning that allows consumers to identify product counterfeiting by means of a mobile phone.

In order to achieve at least one of the above objectives, a method for constructing a machine learning model for ultra-deep regression analysis is proposed as follows.
In order to solve the above-mentioned problems, the invention according to claim 1 is a method for constructing a machine learning model for regression analysis.
(1) To configure a machine learning system for regression analysis by installing software including at least regression analysis and self-organization based on the maximum probability scale on hardware including a processor and memory ;
(2) The machine learning system for regression analysis described above collaborates with hardware and software to create a linear regression line obtained from all data in the processing area and the processing areas on both sides of the linear regression line. Perform self-organization processing based on the maximum probability scale only for the contained data to generate a new processing area;
(3) Regression analysis characterized in that the result of linear regression analysis is obtained by a process of repeating until the difference in data density from the previous iterative process in the processing area generated as described above becomes a predetermined value or less . How to configure a machine learning model.

The probability scale of self-organization processing is normal distribution, multivariate normal distribution, logarithmic normal distribution, exponential distribution, t distribution, F distribution, X2 distribution, ^binomial distribution, negative binomial distribution, polynomial distribution, Poisson distribution, Maximum probability density consisting of probability distributions including at least one of Alan distribution, hypergeometric distribution, geometric distribution, traffic distribution, Weber distribution, triangular distribution, beta distribution, gamma distribution, probability distribution dispersion, mean square deviation, or As a measure of randomness including covariance
The machine learning model for regression analysis according to claim 1, wherein the maximum probability value of the self-organization is an average value or an expected value of a probability distribution obtained by the self-organization based on the scale of the maximum probability value. Is the configuration method of.

The configuration of the machine learning model for regression analysis according to claim 1, wherein the self-organization based on the probability scale transfers the probability space in the maximum probability direction by processing the self-organization based on the probability scale. Method.

The method for constructing a machine learning model for regression analysis according to claim 1, wherein the density value of the data in the processing region is set to the density of the data distribution with respect to the new processing region generated by the repeated processing. Is.

In order to solve the above-mentioned problems, the invention according to claim 5 is a device for machine learning of regression analysis.
(1) A machine learning system for regression analysis is configured by installing software including at least regression analysis and self-organization based on the maximum probability scale on hardware including a processor and memory.
(2) The above-mentioned machine learning system for regression analysis collaborates with hardware and software to obtain a linear regression line from all the data in the processing area and the processing areas on both sides of the linear regression line. Self-organization processing based on the maximum probability scale is performed only on the data contained in, and a new processing area is generated .
(3) Regression analysis characterized in that the result of linear regression analysis is obtained by a process of repeating until the difference in data density from the previous iterative process in the processing area generated as described above becomes a predetermined value or less . Machine learning component .

In order to solve the above problems, the invention according to claim 6 is a program for causing a processor to execute the method according to any one of claims 1 to 4.

In order to solve the above problems, the invention according to claim 7 is a general-purpose mobile terminal device equipped with the program according to claim 6.

The machine learning model of regression analysis submitted by the present invention can collectively determine the regression analysis and the distance from all the data in the processing area to the linear regression-analyzed straight line. In addition, it is possible to become completely autonomous learning by processing self-organization based on a probability scale, which does not require human intervention for all processing. In addition, by performing self-organization processing based on the probability scale, it is possible to generate a new processing area based on the scale of the maximum probability value for the distance between the regression-analyzed straight line and all the data on both sides. It will be possible. In particular, by fusing the role of the regression analysis model with the self-organization model based on the probability scale, both roles can be maximized. From the result of simply performing ordinary regression analysis, it becomes possible to dramatically and accurately recognize the white line recognition used for autonomous driving even in dark weather or in complicated environments, and probabilistic data prediction. It is possible to make dramatic and accurate predictions for the fields of. Expected to be widely applied to industrial artificial intelligence.

As will be described below, embodiments of the present invention will be described in more detail with reference to the accompanying drawings, but embodiments of the present invention are exemplary and not limited.

FIG. 1 is a diagram showing a multi-probability scale definition in a probability distribution.
In FIG. 1, for a Euclidean space including a plurality of probability spaces, the scale value of the probability distribution can be obtained from the value of the probability distribution in each probability space. Here, the center value of the probability distribution is (102), the first scale value is (103), and the region between them is (106). Next, the second scale value is set to (104), and the region between them is set to (107). Further, assuming that the third scale value is (105) and the region between them is (108), the scale value of the probability distribution in a plurality of regions is called a multi-probability scale.

Another way to set up a multi-probability scale is to calculate the spacing between the first tick marks and then make the spacing between the second and third tick marks similar to the spacing between the first tick marks, eg Figure. The interval between 1 (103) and (104) and the interval between (104) and (105) can be the same as the interval between (102) and (103). In addition, the value of the interval of the first scale value is used as the maximum probability scale.

Multi-probability scales, normal distribution, multivariate normal distribution, logarithmic normal distribution, exponential distribution, t distribution, F distribution, X2 distribution, ^binomial distribution, negative binomial distribution, polynomial distribution, Poisson distribution, Arlan distribution (Erlang) From a probability distribution that has at least one probability characteristic for a distribution), a hypergeometric distribution, a geometric distribution, a Weber distribution (Weibull Distribution), a triangular distribution, a beta distribution (Beta is a Distribution), and a gamma distribution (Gamma Distribution). Will be.

The multi-probability scale can be a correlation coefficient for a probability space or a distance belonging to a different space.
In addition, as a non-probability space, the multi-probability scale includes the Euclidean Distance scale, the Manhattan Distance scale, the Chebyshev Distance scale, and the Minkowski Distance (Minkowski Distance) scale. Scale, Mahalanobis Distance scale, Cosine scale, W distance (Wasserstein Distance) scale, KL distance (Kullback-Leibler Distance) scale, PE distance (Pearson Scale) including at least PE distance (PearsonDist) It may consist of one type of scale.

The probability scale may be used as a measure of the maximum probability density, for example, the variance of the probability distribution, the root mean square deviation, or the covariance, and the value of the maximum density of the data distribution in the one-dimensional space, the two-dimensional space, or the multidimensional space. You can do it.

The multi-probability scale may be used as a Jaccard index (Jaccard index) scale, a Hamming Distance scale, or an Information Entropy scale.

The multi-probability scale may be used as a probabilistic algorithm, for example, a multi-probability scale can be defined by using a Bayesian method, a Gaussian process, a Gaussian process and a Bayesian hybrid algorithm, and the like. It is possible.

In order to introduce the contents of the present invention with emphasis, I will not introduce the specific mathematical formulas of the above probability scales one by one, but using a method that will be all the scales for the data, based on the probability scale, It is within the scope of the present invention to construct a machine learning model.

FIG. 2 is a flowchart of self-organizing machine learning based on a multi-probability scale.
Here, as shown in FIG. 2, the flow of the machine learning process of the self-organization is as follows. Here, the position of the scale of the probability distribution and the maximum probability distribution value of the corresponding region are obtained for the objective function. Is.

Initialization step S ₁ : Initialization step of the self-organizing program based on the multi-probability scale. In this step, the input objective function D (x, y) can be either one-dimensional data, two-dimensional data, or arbitrary dimension.

Here, the initial probability scale is M ⁽⁰⁾ , and the initial center value is (x ₀ , y ₀ ) ⁽⁰⁾ . Here, the initial center value is referred to as an initial feature value.

Self-organization based on the multi-probability scale has two processing methods. The first method performs self-organization based on the multi-probability scale from the outside to the center of the probability distribution, and the second method conversely performs self-organization based on the multi-probability scale from the center to the outside of the probability distribution. conduct.

Self-organization initialization method based on the first probability scale:
In this case, the initialization multi-probability scale M ⁽⁰⁾ 'is set as the largest scale for the corresponding probability distribution, and therefore the probability scale is set as M ⁽⁰⁾ '> M ⁽⁰⁾ , generally 3. It will be about double.

Self-organization initialization method based on the second probability scale:
Here, M ⁽⁰⁾ '= M ⁽⁰⁾ may be set, and the M ⁽⁰⁾ value may be artificially predicted without the need for strict setting. Further, the maximum probability scale M ⁽⁰⁾ is set as the radius value, and the circular region or the rectangular region is set as the self-organized region according to the initial probability scale M ⁽⁰⁾ and the center value (x ₀ , y ₀ ) ⁽⁰⁾ . If the maximum probability scale M ⁽⁰⁾ is possible, at least a part of the learned final result M ⁽ⁿ⁾ should be included. If the initialization probability scale M ⁽⁰⁾ 'is too large, the calculation time will be long. On the contrary, if it is too small, accurate results may not be obtained.

For other initialization settings, V is the convergence value of the self-organization, and the difference between the result of the previous organization and the result of the current self-organization is used as a condition for determining whether or not the processing of the self-organization is completed. If the convergence value V is too large, accurate results may not be obtained, but if the convergence value V is too small, the calculation time becomes long. The correct setting method is about 5-10% of the final probability scale M ⁽ⁿ⁾ . The maximum number of repetitions of the self-organization is set to MN, and generally 5 to 10 times in order to avoid an infinite repetition state with respect to the self-organization. Also, let m be the number of scales on the multi-probability scale. Regarding the setting of several meters of the scale of the probability scale, for example: When there are three scales of the multi-probability scale, it can be divided as three scales with respect to the probability distribution area, in which case m = 3. The default number of processes of the own organization is set to n = 0.

Self-organization step S2 based on the multi-probability scale: _In this step, for the objective function D (x, y), which is to perform the nth self-organization processing of the m-th probability scale, (x, y). _i , y _j ) _m ^(n-1) is the center value of the self-organization, the mth probability scale M _m ^(n-1) is the radius, and the center value (x _i , y _j ) _m ^(n-1) And the radius M _m ^(n-1) value, the data of the new objective function _dm ⁽ⁿ⁾ (x _i , y _j ) (i = 1, 2, ..., k; j = 1, 2, ..., L) Can be generated. For the new probability distribution, the m-th new center value ( _xi , y _j ) _m ⁽ⁿ⁾ and the new probability scale M _m ⁽ⁿ⁾ can always be obtained.

Here, dm ( _n ⁾ (x _i , y _j ) ∈ D (x, y), n = n + 1, MN = MN-1. That is, in this step, by generating one new objective function for each objective function data dm ( _n ⁾ ( _xi , y _j ), it is possible to gradually approach the maximum probability direction. be.

を判断する、もし１つの条件を満たすとすれば、多確率尺度のｍ番目の確率尺度に基づいた自己組織が終了する、この場合にＳ_４ステップへ、もしそうではなければ、Ｓ_２ステップへ移行する。続けて多確率尺度に基づいた自己組織の処理を行う。Judgment of processing completion Step S ₃ : Judgment as to whether or not the processing of the mth probability scale has been completed.

If one condition is met, the self-organization based on the m _- th probability scale of the multi _- probability scale ends, in this case go to step S4, if not, go to step S2. Transition. Subsequently, the self-organization is processed based on the multi-probability scale.

次に、第２の確率尺度に基づいた自己組織方法に対し、多確率尺度は最小目盛から始める場合に、多確率尺度を正規分布からなることとし、以降の確率尺度の目盛りは一番目の最大確率尺度目盛値を同じとすると、多確率尺度に基づいた自己組織の処理の必要はない。一番目の最大確率尺度の間隔を他の目盛りの間隔と同様とする。また、確率分布表から推算することも可能である。この場合に、確率尺度に基づいた自己組織の処理方法とする、ｍ＝０、また次の多確率尺度に基づいた自己組織が終了判断ステップＳ_５へ続ける。Data storage step S ₄ : After the self-organization processing based on the m-th probability scale is completed, the feature values of the m-th scale (x _i , y _j ) _m ⁽ⁿ⁾ , the m-th probability scale M _m ⁽ⁿ ). ⁾ Is saved as a result of learning. Here, the feature value (x _i , y _j ) _m ⁽ⁿ⁾ of the scale is used as the probability distribution value in the m-th region, and the probability scale M _m ⁽ⁿ⁾ is used as the m-th scale. Next, m = m-1, and the number of scales of the probability scale is corrected. With respect to the set of objective functions, that is, the data range of the self-organization, the new probability distribution is as follows by the multi-probability scale M _m ⁽ⁿ⁾ and the feature values ( _xi , y _j ) _m ⁽ⁿ⁾ .

Next, in contrast to the self-organization method based on the second probability scale, when the multi-probability scale starts from the minimum scale, the multi-probability scale consists of a normal distribution, and the subsequent scales of the probability scale are the maximum of the first scale. If the probability scale scale values are the same, there is no need to process the self-organization based on the multi-probability scale. The interval of the first maximum probability scale is the same as the interval of other scales. It is also possible to estimate from the probability distribution table. In this case, m = 0, which is the processing method of the self-organization based on the probability scale, and the self-organization based on the next multi _- probability scale continues to the end determination step S5.

End determination step S ₅ : It is a determination step as to whether or not the processing of the self-organization based on the multi-probability scale is completed. In this step, if m = 0, the processing of the self-organization based on the multi-probability scale is terminated. _In this case, the process proceeds to step _S6 , and if not, the process proceeds to step S2.
Return step S ₆ : Finally finish and return to the main program.

Next, the distance between one point in Euclidean space and the center point of the probability distribution in probability space is defined.
FIG. 3 is a diagram showing a unified definition of distance between Euclidean space and probability space.
As shown in FIG. 3, (301) is a Euclidean space including the probability space, (302) is the center point of the probability distribution in the probability space w _j ∈ W, and (303) is the multi-probability of the probability distribution in the probability space. Let M ₁ be the first scale of the scale, (304) be the second scale M ₂ of the multi-probability scale of the probability distribution in the probability space, and (305) be the third scale of the multi-probability scale of the probability distribution in the probability space. Let M ₃ be the scale of, and let (309) be the point v _j ∈ V in the Euclidean space. Find the distance from the set V to the set W belonging to the probability space.

ここに，

In actual pattern recognition, the data of each obtained feature value is random under different conditions. By performing iterative learning, it is possible to obtain the value of the probability distribution of n feature values in which the feature vector is composed.
Therefore, let w _j ∈ W be one point (302) in the set W belonging to the probability space, and let v _j ∈ V be one point (309) in the set V. Further, when (j = 1, 2, ..., N) is used as the counter value of the feature value, the interval between (302) and (303) is set to D _1j ^(wj) = M with respect to the counter value j of the feature value. _2j −M _1j , and the probability distribution value in this region is p _1j ^(wj) . The interval between (303) and (304) is D _2j ^(wj) = M _3j −M _2j , and the probability distribution value in this region is p _2j ^(wj) . The interval between (304) and (305) is D _3j ^(wj) = M _4j − M _3j , and the probability distribution value in this region is p _3j ^(wj) . As shown in FIG. 3, the number of scales is m _j ^(wj) = 3 for the region of three scales of the probability distribution belonging to w _j from V to W. The distance G (V, W) in which the Euclidean space and the probability space are unified between (309) and (302) can be defined as follows.

Here,

Δ _j ^(wj) in the above equation is an error value between different spatial distances. Therefore, if this error value is corrected, it becomes possible to obtain a distance that unifies different spaces. Here, the distances in different spaces are the Euclidean distance and the probability space distance in the space of probability. That is, it is the error between the distance value obtained by the Euclidean distance formula and the actual distance value in the probability space in the space of probability. According to the above concept, if Δ _j ^(wj) is used as a correction value and correction is made, it becomes possible to accurately obtain the distance at which the Euclidean space and the probability space are unified. For the machine learning domain, it has become possible to solve the scale problem between data to which Euclidean space and probability space belong to both.

If one point r _j ∈ R belonging to the set R exists in the interval from v _j to w _j , then the formula representing the ambiguous relationship between r _j (310) and w _j (302). Is as follows.

Let the above equation be a membership function. The closer an arbitrary point r _j ∈ R is to w _j ∈ W, the closer the result of F _j ^(wj) is to 1, and conversely, the farther r _j ∈ R is from w _j ∈ W, the more F The result of _j ^(wj) approaches 0. Here, since the method of expressing the distance consists of the number 6, the result of ^{Fj (wj)} _is merely ambiguous information in which the Euclidean space and the probability space are unified.

The method of defining the membership function is not limited to the method provided above, but various formulas can be defined by the method of human intervention, so no matter how it is defined, it is between the two elements of the objective function. If we define an ambiguous relationship, they all fall within the scope of the present invention.

Here, if any one point r _j ∈ R, and if it happens to be in the i-th region D _ij ^(wj) of the probability distribution of w _j ∈ W, the probability distribution value of this place is pf _j ^(wj) , and also. , Another possibility, if v _j ∈ V also happens to be in the qth region D _qj ^(wj) of the probability distribution of w _j ∈ W, then the probability distribution value at this location is ph _j ⁽ Let ^wj) , and the formula of the fuzzy event probability measure F ^(w) between the set R and the set W is as follows.

According to the above definition, the role of the fuzzy event probability measure F ^(w) is fairly stable information at the macro level by integral calculation using minute ambiguous information and minute unstable probability information. Is only the theoretically optimal method of information processing for the criteria in the strictest similarity relationship between the R and W of the set, which is similar to the two sets in the application of pattern recognition. Since it is possible to reflect the relationship to the maximum extent, it is possible to optimally collate the feature vector set R with the registered dictionary data set W having probability distribution information.

As shown in Eq. 8, r _j ∈ R and v _j ∈ V set the probability distribution value to pf _j ^(wj) = 0 and ph _j if they are not included in the probability distribution of w _j ∈ W. ^{When (wj)} = 0, it becomes an ambiguous expression of a membership function related to a vector distance that unifies the utrid space and the probability space, similar to the result of integrating the above number 7.

The above Euclidean space is referred to as Manhattan space: (Manhattan Space); Haarana Space; Chebyshev Space; Minkowski Space; Malpigi space (Mahalanobis Space); Can be expanded.

By summarizing the two formula methods in which the Euclidean space and the stochastic space distance are unified by the equations 7 and 8, it was found that the probabilistic space is included in the Euclidean space, respectively. When the data reaches a region of probability space, the value of that distance is related to the probability distribution value of each region passed.

In addition, the idea so far is that the distance scale of the probability space is related to the probability distribution value of the passed region when passing through the probability space, and it is necessary to have a certain direction, which is general. Unlike the distance scale, it does not meet the condition of symmetry. For example, when calculating the distance from v _j to w _j , the distance in the probability space is related to the value of the probability distribution of w _j in the process of passing from the position of v _j to the position of w _j . Even if it has a probability distribution of v _j , it has nothing to do with the probability distribution value of v _j . This does not satisfy the symmetry of the probability scale and the triangle inequality. However, the present invention also submits a method capable of satisfying all the conditions for measuring the probability as described below.

FIG. 4 is a schematic diagram of a pattern recognition model for ultra-deep competitive learning.
In ultra-deep competitive learning, as shown in FIG. 4: Two feature vector data fv _1j ∈ FV ₁ and fv _2j ∈ whose maximum probability is obtained by processing (4100) and (4200) by self-organization based on the probability scale. Let FV ₂ be, and let (4000) be a feature element sv _j ∈ SV (i = 1, 2, ... e) belonging to the feature vector SV to be recognized. Further, (4001) is sv ₁ , (4002) is sv ₂ , (4003) is sv ₃ , ..., (400e) is sv _e .

Feature vector data The first scale of the feature element fv ₁₁ belonging to FV ₁ is (4111), the second scale of fv ₁₁ is (4112), the third scale of fv ₁₁ is (4113), and fv. Let the center value of ₁₁ be (4110). Further, the first scale of fv ₁₂ is (4121), the second scale of fv ₁₂ is (4122), the third scale of fv ₁₂ is (4123), and the center value of fv ₁₂ is (4120). And. Similarly, the first scale of fv ₁₃ is (4123), the second scale of fv ₁₃ is (4132), the third scale of fv ₁₃ is (4133), and the center value of fv ₁₃ is (4130). ). Further, the first scale of fv _1e is (41e1), the second scale of fv _1e is (41e2), the third scale of fv _1e is (41e3), and the center value of fv _1e is (41e0). And.

The first scale of the feature element belonging to the feature vector data FV ₂ is (4211), the second scale of fv ₂₁ is (4212), and the third scale of fv ₂₁ is ₍ 4213). Let the center value of fv ₂₁ be (4210). Further, the first scale of fv ₂₂ is (4221), the second scale of fv ₂₂ is (4222), the third scale of fv ₂₂ is (4223), and the center value of fv ₂₂ is (4220). And. Similarly, the first scale of fv ₂₃ is (4231), the second scale of fv ₂₃ is (4322), the third scale of fv ₂₃ is (4233), and the center value of fv ₂₃ is (4230). And. Further, the first scale of fv _2e is (42e1), the second scale of fv _2e is (42e2), the third scale of fv _2e is (42e3), and the center value of fv _2e is (42e0). And.

For the feature element sv _j ∈ SV belonging to the feature vector SV to be identified, if it is in the region of the feature element fv _2j probability distribution belonging to the feature vector FV ₂ registered by chance, the probability distribution value in that region is set. Let sf _j ^(fv2j) ∈ SF ⁽ FV2). Further, when the center value of the probability distribution of fv _1j happens to be in the area of the probability distribution of the feature element fv _2j , the probability distribution value is sh _j ^(fv2j) ∈ SH ^(FV2) (j = 1, 2, ... , E). (This is a case where a very special situation occurs, that is, the probability distribution of fv _1j and the probability distribution of fv _2j almost overlap.)

The scale interval of the multi-probability scale of the probability distribution of fv _2j passing from sv _j ∈ SV to the center of the probability distribution of fv _2j is ^{Dij (fv2j)} _, the number of multi-probability scales is mj ( _fv2j ⁾ , and _Dij . Let the probability distribution value of fv _2j in the interval of ^(fv2j) be Pij ( _fv2j ⁾ (i = 1, 2, ..., Mj ( _fv2j ⁾ ).

Based on the equations 7 and 8, the fuzzy event probability measure for the feature vector SV to be recognized and the registered feature vector data FV ₂ is as follows.

Similarly, for the feature element sv _j ∈ SV belonging to the feature vector SV to be identified, if it is in the probability distribution region of the feature element fv _1j belonging to the feature vector FV ₁ registered by chance, the probability distribution value thereof. Is sf _j ^(fv1j) ∈ SF ^(FV1) , and when the center value of the probability distribution of fv _2j happens to be in the probability distribution region of the feature element fv _1j , the probability distribution value is sh _j ^(fv1j). Let ∈ SH ^(FV1) (j = 1, 2, ..., E). (This is for a very special situation, that is, the probability distribution of fv _2j and the probability distribution of fv _1j almost overlap.)

When passing from sv _j ∈ SV to the center of the probability distribution of fv _1j , the scale interval of the multi-probability scale of the probability distribution of fv _1j is ^{Dij (fv1j)} _, and the number of multi-probability scales is mj ( _fv1j ⁾ . , The probability distribution value of fv _1j in the region of D _ij ^(fv1j) _is ^{Pij (fv1j)} (i = 1,2, ..., Mj ( _fv1j ⁾ ). The fuzzy event probability measure for the feature vector SV to be recognized and the registered feature vector data FV ₁ is as follows:

或は

ここで、Ｆ＞１になると、識別対象になる特徴ベクトルＳＶは特徴ベクトルデータＦＶ_２に属する、逆にＦ＜１になると、識別対象になる特徴ベクトルＳＶは特徴ベクトルデータＦＶ_１に属する。The feature vector SV to be identified competes with the registered feature vector data FV ₂ and FV _1. The formula is as follows:

Or

Here, when F> 1, the feature vector SV to be identified belongs to the feature vector data FV ₂ , and conversely, when F <1, the feature vector SV to be identified belongs to the feature vector data FV ₁ .

FIG. 5 Schematic diagram of an optimal classification model using ultra-deep competitive learning As shown in FIG. 5,: (501) is defined as an Euclidean space including a probability space. Further, in the Euclidean space (501), there are probability distributions (520) and (530) of two probability spaces. Further, (502) is set as the center value of the probability distribution (520). (503) is the first scale of the probability distribution (520), (504) is the second scale of the probability distribution (520), and (505) is the third scale of the probability distribution (520). Next, (506) is defined as the first scale region of the probability distribution (520), and the probability distribution value in this region is defined as _p1j ⁽⁵²⁰⁾ . Further, (507) is defined as the second scale region of the probability distribution (520), and the probability distribution value in this region is defined as p _2j ⁽⁵²⁰⁾ . Further, (508) is defined as the third scale region of the probability distribution (520), and the probability distribution value in this region is defined as p _3j ⁽⁵²⁰⁾ .

Similarly, (510) is the center value of the probability distribution (530), (511) is the first scale of the probability distribution (530), and (512) is the second scale of the probability distribution (530). 513) is the third scale of the probability distribution (530), (514) is the first scale region of the probability distribution (530), and the probability distribution values in this region are _p1j ⁽⁵³⁰⁾ and (515) are probabilities. The second scale region of the distribution (530), the probability distribution value of this region is p _2j ⁽⁵³⁰⁾ , (516) is the third scale region of the probability distribution (530), and the probability distribution value of this region is Let p _3j ⁽⁵³⁰⁾ .

Next, let the two medians (502) and (510) corresponding to the probability distributions (520) and (530) be the elements w _j ∈ W and v _j ∈ V of the two sets. There is an arbitrary point r _j ∈ R on the straight line connected on both sides of the two (502) and (510). Arbitrary data is a point (500) projected onto this straight line. The optimal classification using ultra-deep competitive learning is to find out which probability distribution 520 or 530 the arbitrary single point r _j ∈ R belongs to.

Here, m _j ^(wj) is the number of probability scales between r _j and the center w _j of the probability distribution, and m _j ^(vj) is the probability between r _j and the center v _j of the probability distribution. It is the number of scales. For example, in FIG. 5, ^{mi (wj)} ₌ 3, _pij ^(wj) = _pij ⁽⁵²⁰⁾ , _pij ^(vj) = _pij ⁽⁵³⁰⁾ [i = 1, 2, ..., ( ^{Mj (wj} ₎ ). ⁾ = M _j ^(vj) )].
According to the equation 6, the expression formula of the exact distance between the set V belonging to the probability space (530) and the set W belonging to the probability space (520) is as follows.

また、

ここで、ｒ_ｊから確率分布の中心ｗ_ｊまでの距離の定式と、ｒ_ｊから確率分布の中心ｖ_ｊまでの距離の定式を加えて、確率分布（５３０）の中心ｖ_ｊから確率分布（５２０）の中心ｗ_ｊまでの距離の定式になることを考えていると、二つの確率分布の値に関係するべきである。そのため、（Δ_ｊ ^（ｖｊ）＋Δ_ｊ ^（ｗｊ））を、確率空間の中で、ユークリッド距離と、確率空間の距離と間での誤差とする。さらに、（Δ_ｊ ^（ｖｊ）＋Δ_ｊ ^（ｗｊ）を、修正値として修正すれば、ユークリッド空間距離と確率空間が統一された厳密な距離を正確的に得ることが可能になる。上記の数５－８の距離の定義は従来の距離尺度条件の対称性と三角不等式を満たさない。これに比べると、数１２は距離の対称性と三角不等式を含め、すべての距離の尺度条件が満たされた、ユークリッド空間と確率空間が統一された距離に対して最も大きな問題を解決した。
上述数８と同様に、ｒ_ｊ∈Ｒと確率分布（５３０）中心値ｖ_ｊ∈Ｖとの間のファジィ事象確率測度を考える時に、もしｒ_ｊ∈ＥＲが（５３０）の確率分布の領域にある場合に、その領域の確率分布値をｐｆ_ｊ ^（ｖｊ）とする。また、ｗ_ｊ∈Ｗが偶然に（５３０）確率分布の領域にある場合に、その領域の確率分布値をｐｈ_ｊ ^（ｖｊ）とする。上記の確率空間の距離の対称性により、もしｖ_ｊ∈Ｖも偶然に（５２０）確率分布の領域にある場合に、その領域の確率分布値をｐｈ_ｊ ^（ｗｊ）とする。（二つの確率分布がほぼ重なった特殊場合とする）。here,

again,

Here, by adding the formula for the distance from r _j to the center w _j of the probability distribution and the formula for the distance from r _j to the center v _j of the probability distribution, the probability distribution (from the center v _j of the probability distribution (530)) Considering that it becomes a formula of the distance to the center _wj of 520), it should be related to the values of the two probability distributions. Therefore, (Δ _j ^(vj) + Δ _j ^(wj) ) is set as the error between the Euclidean distance and the distance in the probability space in the probability space. Further, if (Δ _j ^(vj) + Δ _j ^(wj) is corrected as a correction value, it is possible to accurately obtain an exact distance in which the Euclidean space distance and the probability space are unified. The definition of distance of -8 does not satisfy the symmetry and triangular inequality of the conventional distance scale condition. By comparison, the equation 12 satisfies all the distance scale conditions including the distance symmetry and the triangular inequality. , Solved the biggest problem for the unified distance between Euclidean space and probability space.
Similar to Eq. 8, when considering the fuzzy event probability measure between r _j ∈ R and the probability distribution (530) center value v _j ∈ V, if r _j ∈ ER is in the region of the probability distribution (530). In a certain case, the probability distribution value in that region is pf _j ^(vj) . Further, when w _j ∈ W happens to be in the region of the (530) probability distribution, the probability distribution value in that region is set to ph _j ^(vj) . Due to the symmetry of the distance in the probability space described above, if v _j ∈ V also happens to be in the region of the (520) probability distribution, the probability distribution value in that region is defined as ph _j ^(wj) . (Suppose it is a special case where two probability distributions almost overlap).

ここで、数６により、

また、

By equation 12, any set R can define the formula of the fuzzy event probability measure belonging to the set V as follows:

Here, by number 6,

again,

ここで、数６により、

また、

With reference to the numbers 12 and 13, D _ij ^(wj) and D _ij ^(vj) , p _ij ^(wj) and p _ij ^(vj) , m _j ^(wj) and m _j ^(vj) , pf _j ⁽ Since it is possible to obtain ^vj) and pf _j ^(wj) , ph _j ^(vj) and ph _j ^(wj) , the formula of the fuzzy event probability measure belonging to the set W is as follows.

Here, by number 6,

again,

Finally, with reference to the above number 11, by ^{FF (W)} and F ^(V) , by ultra-deep competitive learning of F = (F ^(W) / F ^(V) ), for the arbitrary set R, It has become possible to classify optimizations between two data with a probability distribution.
Here, the number 12 satisfies the scale conditions of all distances including the symmetry and the triangle inequality for the numbers 13 and 14, so that the fuzzy event probability measure based on this distance is also the symmetry and the triangle inequality. It is now possible to meet all scale conditions.

Here, a model construction method for ultra-deep competitive learning has been described using two probability distributions as an example. However, depending on the actual application, n cases can be classified by either three probability distributions or four probability distributions. Even in the probability distribution, it is possible to classify the optimization using the ultra-deep competitive learning model.

FIG. 6 is a flowchart of an optimum classification model using ultra-deep competitive learning.
As shown in FIG. 6, referring to FIG. 5, ultra-deep competitive learning can be realized by the following eight steps.
_In the initialization step of S1, the content of the initialization of the self _- organization based on the probability scale is the same as that of the initialization step of S1 in FIG. First, let the number of scales of the probability scale be m _j ^(wj) and m _j ^(vj) . For example, m _j ^(wj) = m _j ^(vj) = 3, a data registration space is set, and other necessary initialization processing contents are set.

Steps of self-organization based on the multi-probability scale of S ₂ : See step S2 of FIG. ₂ , for w _jh ∈ Wh and v _jh ∈ V _h ( _h = 1, 2, ..., G). By performing self-organization based on the probability scale, the center values w _j ∈ W and v _j ∈ V of the maximum probability distribution of each element, and the scale values D _ij ^(vj) and D _ij ^(wj ) that became the maximum probability scale. ⁾ And the probability distribution values _pij ^(vj) and _pij ^(wj) of the corresponding region, or when w _j ∈ W and v _j ∈ V and r _j ∈ R are in the probability distribution region, the probability distribution. Find the values pf _j ^(vj) , ph _j ^(vj) and ph _j ^(wj) .

_In the judgment step of S3, it is the end judgment of the self-organization based on the probability scale of S2, and all the data are obtained for the number PN of the probability distribution? " _N " moves to S2, and "Y" moves to step _S4 of data storage.

_In the data storage step of S4, all the data obtained by the step of S2 _are registered as a data pace.

S5 _Did you complete the processing of all probability distributions in the probability distribution judgment step? "N" is the number of probability distributions PN = PN + ₁ , the new probability distribution is calculated in step S2, and "Y" is the next _S6 ultra-deep competitive learning step.

或は

ここで、数６により

また、

２０）に属する、逆の場合に別の確率分布（５３０）に属する。
超深層競合学習の効果は、ユークリッド空間と確率空間が統一された厳密な距離に基づいてファジィ事象確率測度を導入し、ｗ_ｊ∈Ｗ、ｖ_ｊ∈Ｖ及びｒ_ｊ∈Ｒのデータ間で、微小の曖昧な情報と微小の不安定な確率情報を利用して、マイクロレベルで、互いに競合させる、積分計算により、マクロレベル上で最適化されたデータ間での類似関係情報を得ることが可能になった。 _In the S6 ultra-deep competitive learning step, the following ultra-deep competitive learning is performed based on the calculation results of the above equations 13 and 14.

Or

Here, by number 6

again,

It belongs to another probability distribution (530), which belongs to 20), and vice versa.
The effect of ultra-deep competitive learning introduces a fuzzy event probability measure based on the exact distance between the Euclidean space and the probability space, between the data of w _j ∈ W, v _j ∈ V and r _j ∈ R. It is possible to obtain similarity information between data optimized at the macro level by integral calculation that competes with each other at the micro level by using minute ambiguous information and minute unstable probability information. Became.

Was the above ultra _- deep competitive learning completed in the S7 completion judgment step? “N” advances to step _S6 , and ultra-deep competitive learning is performed again. When “Y” is reached, the process proceeds to the next return step.
S8 _In the return step, the process proceeds to the step of returning to the main program.

FIG. 7 is a schematic diagram of a neural network model for ultra-deep competitive learning.
The configuration of ultra-deep competitive learning will be described using image recognition as an example. As shown in FIG. 7, (801) shows a spatial map of the image to be sensed. Ultra-deep competitive learning emphasizes the extraction of image information. In order to maximize the accuracy of image recognition, spatial mapping of various images is performed on the original image to be recognized. For example, there are weekly wave number spatial mapping of an image, color spatial mapping of an image, power spatial mapping of an image, spatial mapping of an edge of an image, and the like. (802) and (803) are shown by taking spatial mappings of two types of images as examples.
(804) is a local region of the image (801). Here, in the case of image identification for mobile phones, a method of dividing a linked image can be used so that the image can be accurately identified at an arbitrary angle of 360 degrees. Introduced. Regarding the method of determining the size of this linked area, the larger the area, the faster the calculation speed, but the recognition accuracy is affected. Conversely, the smaller the area, the higher the recognition accuracy. The calculation speed is relatively slow. The number of pixels and area of each area should be well balanced according to the actual application.

Let (805) be a machine learning module called self-organization based on a plurality of probability scales. A machine learning module called self-organization based on these probability scales is connected between the object of recognition and each contact point of the sensing layer. It has the function of deep excavation for the target function information. It also has the feature of extracting the information of the objective function as the maximum probability. It also has an excellent feature that it can be tracked autonomously according to the misalignment of the image to be recognized.
(806) is used as a sensing layer of a new neural network.

Using (807) as the contact point of the sensing layer of the neural network, the present invention is a method of deepening the function of ultra-deep competitive learning by increasing the amount of information of the target function to contact the contact layer of the sensing layer of the neural network. The number of neural networks and the number of machine learning will be increased accordingly.

Machine learning called self-organization based on a probability scale has the characteristic of autonomously tracking to the feature position of the maximum probability. The problem of image misalignment during image recognition can be improved. It is very important to have this feature, especially when recognizing video.

Let (808) be a machine learning module called self-organization based on a connected probability scale between the sensing layer and the nerve layer. It has a deep excavation function for the probability information input from the sensing layer mainly by ultra-deep competitive learning. Furthermore, it is possible to obtain information on the probability distribution of the objective function. All the learned data is registered as a (809) database.

(810) is the neural layer of the neural network, (811) is the contact point of the neural layer of the neural network, and (814) is the cerebral skin layer of the neural network. Here, the data of the probability distribution information obtained by the ultra-deep competitive learning is transmitted from the nerve layer to the cerebral skin layer.

(812) is an ultra-deep competitive machine learning module connected between the nerve layer (810) and the cerebral cortex (814), and its important function is to make a final decision mainly by the cerebral cortex (814). Is as follows.
Using the above numbers 12 and 16, based on a strict distance scale that unifies the Euclidean space and the probability space, and the fuzzy event probability measure, the sample data to be identified is a plurality of registered probability distribution information. Find the result of ultra-deep competitive learning between feature vector data. The effect of this processing is the effect of weighting the optimization on the elements of the vector of different values as pattern recognition, which can be used for the optimization of both the ambiguous information and the probability information to the maximum extent.

Next, based on the value of the maximum probability scale consisting of ultra-deep competitive learning and the competition result consisting of ultra-deep competitive learning such as equations 11 and 15, it is possible to obtain the threshold value of the nerve layer, which is the final. The final recognition result can be obtained from the cerebral cortex, which is a threshold value for nerves in the cerebral cortex (814), which is a processing method that mimics brain function.

Finally, for each registered feature vector data, the final recognition result is determined using the feature vector data with the maximum reliability, which is machine learning of self-organization based on the probability scale in reliability.

Specifically, the method for obtaining the reliability of the feature vector is to set the success rate to CD _kj and the erroneous recognition rate based on the result of actually recognizing the j-th feature element of the feature vector of the k-th image already registered. Is ED _kj , and the reliability value can be expressed by the following number 17.

Next, in order to obtain the maximum probability reliability, and for the reliability value of the feature element of each feature vector of each of the above images, machine learning called self-organization based on the probability scale is performed to obtain the maximum probability. A feature element with the maximum probability reliability is selected using the probability scale as a criterion, and only the feature element with the maximum probability reliability is subjected to ultra-deep learning. As a result, the identification result with the maximum reliability can be obtained.

As described above, the feature value information of the maximum probability of the target image is transmitted to the sensing layer by machine learning called self-organization based on the probability scale. From the sensing layer, the probability distribution information obtained by machine learning called self-organization based on the probability scale is transmitted to the nerve layer. From the nerve layer, the value of the reliability of the maximum probability and the scale value of the maximum probability obtained by machine learning of self-organization based on the probability scale are transmitted to the cerebral skin layer. As described above, the feature of the new neural network is to transmit probability information to each layer, and to connect to a machine learning module called self-organization based on a probability scale between the contacts of each layer. ..

FIG. 8 is a schematic diagram of the introduction of ultra-deep competitive learning in the identification of counterfeit products in mobile phones.
As shown in FIG. 8, the two color spaces of the electronic image and the printed image, in which the print image color is a color space composed of CMYK and the electronic image color is a color space composed of RGB, almost overlap each other. However, there are some parts that do not overlap with each other, and it is possible to create a mechanism for recognizing counterfeit products with a mobile phone by utilizing such characteristics.

Here, if the color of the original image is (1101) and the scanned image is (1102), as shown in FIG. 8, the scanned printed image is not exactly the same as the original image.
However, with some modification methods, the printed image copied by scanning can be made closer to the original printed image. It is difficult to clearly distinguish primitive images, especially with the naked eye and with conventional optical classifiers. It is difficult to determine which is the copied image, which is a difficult product anti-counterfeiting problem that cannot be solved by the entire society today.

A specific method is to use an optical classifier to identify multiple original print images on a print line, or to take multiple original print images on a mobile phone under different environments. Using machine learning of self-organization based on the multi-probability scale, the probability information of the feature vector of the original printed image is obtained. Here, the probability information includes the feature value of the maximum probability distribution, the value of the maximum probability scale, and the value of the maximum probability scale. It is to include the maximum probability distribution value of each feature element.

Next, in the process of identifying the fake image on the mobile phone, the data of each feature element of the image recognized by the mobile phone is maximized by self-organizing machine learning based on the multi-probability scale, as shown in FIG. After obtaining the characteristic value of the probability, it is input to each contact point of the sensing layer, the registered data and the ultra-deep counter-learning are performed between the sensing layer and the nerve layer, and the value of the probability measure of the fuzzy event is acquired. After machine learning of the self-tissue based on the multi-probability scale, the reliability of the nerve layer and the cerebral cortex is finally determined whether the printed image read by the cerebral cortex is a counterfeit product.

Fig. 9 Process flow chart of machine learning of ultra-deep linear regression analysis As shown in Fig. 9, the process flow of ultra-deep linear regression analysis learning consists of the following 6 steps:
Initialization step S ₁ : First, the target data of the ultra-deep linear regression analysis learning is the data belonging to the data set RD ⁽ⁱ⁾ in the i-th processing area a ⁽ⁱ⁾ (x j (i), y j (x _j ⁽ⁱ⁾ , y _j ^{( i)} ) ∈ RD ⁽ⁱ⁾ (j = 1, 2, ..., q; i = 1, 2, ..., N _max -1), and the data density in the i-th processing area a ⁽ⁱ⁾ is D _ns ^{( i)} . Here, i is a counter value of the number of times of ultra-deep linear regression analysis learning, and N _max is the maximum number of times of ultra-deep linear regression analysis learning. The initial values are (x _j ⁽⁰⁾ , y _j ⁽⁰⁾ ) ∈ RD ⁽⁰⁾ , D _ns ⁽⁰⁾ , the processing effect judgment value V _eff , and the initial processing area a ⁽⁰⁾ .

Calculation straight line distance step S ₂ : The calculation of the straight line subjected to regression analysis for the data (x _j ⁽ⁱ⁾ , y _j ⁽ⁱ⁾ ) ∈ RD ⁽ⁱ⁾ in the processing area a ⁽ⁱ⁾ is as follows. :

Here, the result of the i-th conventional linear regression analysis is that y ⁽ⁱ⁾ 'is the mean value of y ⁽ⁱ⁾ , x ⁽ⁱ⁾ ' is the mean value of x ⁽ⁱ⁾ , and b ⁽ⁱ ). ⁾ Is the slope value of the linear regression analysis. In the processing area a ⁽ⁱ⁾ , the distance from the j-th data (x _j ⁽ⁱ⁾ , y _j ⁽ⁱ⁾ ) ∈ RD ⁽ⁱ⁾ (j = 1, 2, ..., Q) to the linear regression-analyzed straight line is It looks like this:

加する量、或は減少する量θの値は、黄金分割探索方法により決定すれば、最小繰り返し回数として超深層線形回帰分析学習の処理を行うことが可能である。Self _- organization step S3 based on the probability scale: From (x _j ⁽ⁱ⁾ , y _j ⁽ⁱ⁾ ) ∈ RD ⁽ⁱ⁾ based on the processing of the self-organization based on the probability scale in FIGS. 1 and 2 above. Regression analysis Self-organization processing based on the probability scale is performed on q _dj ⁽ i), where the distance to the linear distance is _dj ⁽ⁱ⁾ (j = 1, 2, ..., Q).
New region generation step S4, with the regression diagonal line performed i times of linear regression analysis as the center line based on the maximum probability value M ₍ ⁱ⁾ obtained by processing the self-organization based on the probability scale in the above step. If data within the maximum probability value is reserved on both sides and data other than the maximum probability value is removed, a new target data set RD ^{(i + 1)} for learning ultra-deep linear regression analysis in the new region a ⁽ⁱ + 1) is generated. ..
Here, the data density in the i-th processing area a ⁽ⁱ⁾ is D _ns ⁽ⁱ⁾ , which is the density of the data in the corresponding processing area, and an example is an image of 300 pixels in a 20 * 30 pixel area. Yes, the data density value is D _ns ⁽ⁱ⁾ = 300 / (20 * 30) = 0.5

If the value of the amount to be added or the amount to be decreased θ is determined by the golden section search method, it is possible to perform the processing of ultra-deep linear regression analysis learning as the minimum number of repetitions.

また、

数２４により、もし「Ｎ」になると、Ｓ_２に移行する、その他「Ｙ」になると、次の戻るステップＳ_６へ移行する。End judgment step S ₅ : Next is a method of judging whether or not the processing of ultra-deep linear regression analysis learning has been completed.

again,

According to the number 24, if it becomes "N", it shifts to S2, and if it becomes "Y", it shifts to the _next return step _S6 .

Back Step S ₆ : The process of ultra-deep linear regression analysis learning is completed, and the process returns to the main program.

The neural network for ultra-deep learning submitted in the present invention has a special function of performing distributed processing by machine learning for each important part of the system. It is possible to apply industrial AI systems such as autonomous driving. In addition, a control system with a new automatic control method can be expected. Furthermore, since the number of points of contact between the sensing layer and the nerve layer, or the number of machine learning, can be increased infinitely, the performance of machine learning can be increased infinitely accordingly. Furthermore, it is possible to select whether it is small or large for the hardware environment, depending on the application. It is no exaggeration to say that these characteristics make it possible to apply the AI system to all fields of industry.

Diagram showing the definition of a multi-probability scale in a probability distribution Flowchart diagram of machine learning of self-organization based on multi-probability scale Diagram showing the definition of a unified distance between Euclidean space and probability space Diagram showing a pattern recognition model using ultra-deep competitive learning Schematic diagram of the optimal classification model using ultra-deep competitive learning Flow chart of optimal classification model processing using ultra-deep competitive learning Schematic diagram of neural network model for ultra-deep competitive learning Schematic diagram of the introduction of ultra-deep competitive learning in the identification of counterfeit products on mobile phones Processing flowchart of ultra-deep linear regression analysis learning

A brief description of the sign

101 Probability distribution in probability space 102 Center value of probability distribution 103 First scale value 104 Second scale value 105 Third scale value 106 Area of first scale value 103 107 Between second scale values 103 and 104 Region 108 Region between third scale values 104 and 105 302 Center point of probability distribution in probability space W _j
303 First scale of the multi-probability scale of probability distribution in probability space M ₁
304 Second scale M ₂ of the multi-probability scale of probability distribution in probability space
305 Third scale M ₃ of the multi-probability scale of probability distribution in probability space
309 Elements in Euclidean space v _j
310 Any element r _j ∈ R (j = 1, 2, ..., N) between the element v _j ∈ V in Euclidean space and the center value w _j ∈ W of the probability distribution.
4000 Identification image vector 4001 Identification image feature sv ₁ element 4002 Identification image feature sv ₂ element 4003 Identification image feature sv ₃ element 400e Identification image feature sv _e element 4100 Self-organized feature vector data based on multi-probability scale FV ₁
4200 Self-organized feature vector data FV ₂ based on multi-probability scale
4110 fv ₁₁ center value 4111 feature vector set FV ₁ first scale 4112 fv ₁₁ second scale 4113 fv ₁₁ third scale 4120 fv ₁₂ center value 4121 _{fv 12 first} 4122 fv ₁₂ second scale 4123 fv ₁₂ third scale 4130 fv ₁₃ center value 4231 fv ₁₃ first scale 4232 fv ₁₃ second scale 4233 fv ₁₃ third scale 41e0 fv _1e center value 41e1 fv _1e first scale 41e2 fv _{1e second} scale 41e3 fv _1e third scale 4210 fv ₂₁ center value 4211 feature vector set FV ₂₁ first scale 4212 2nd scale of fv ₂₁ 4213 3rd scale of fv 4220 center value of fv ₂₂ 4221 fv ₂₂ 1st scale 4222 fv ₂₂ 2nd scale 4223 fv ₂₂ 3rd scale 4230 fv ₂₃ center value 4231 fv ₂₃ first scale 4232 fv ₂₃ second scale 4233 fv ₂₃ third scale 42e0 fv _2e center value 42e1 fv _2e first scale 42e2 fv _2e second scale 42e3 fv _2e The element r _j ∈ R (j = 1, 2, ..., N) on the straight line connecting the centers of the two probability distributions of the third scale 500 502 and 510.
501 Euclidean space containing probability space 502 Center value of probability distribution 520 503 First scale of probability distribution 520 504 Second scale of probability distribution 520 Third scale of probability distribution 520 506 First scale of probability distribution 520 The probability distribution value in the area is p _1j
507 Probability distribution The probability distribution value in the first scale area of 520 is p _2j .
508 Probability distribution The probability distribution value in the first scale area of 520 is p _3j .
510 Center value of probability distribution 530 511 First scale of probability distribution 530 512 Second scale of probability distribution 530 513 Third scale of probability distribution 530 514 Probability distribution value p in the first scale region of probability distribution 530 _1j
515 The probability distribution value in the second scale region of the probability distribution 530 is p _2j .
516 The probability distribution value in the third scale region of the probability distribution 530 is p _3j .
Euclidean space including 501 520 and 530 probability spaces 500 An element on a straight line connecting the centers of two probability distributions 502 and 510 r _j ∈ R (j = 1, 2, ..., N)
801 Spatial mapping of the image to be sensed 802 and 803 Spatial mapping of two types of images 804 Local area of image 801 805 A machine called a self-organization based on a multi-probability scale connected between the sensing target and the sensing layer. Learning unit 806 New neural network sensing layer 807 New neural network sensing layer contacts 808 Self-organized machine learning unit 809 database 810 based on a probabilistic scale with ultra-deep competitive learning capabilities connected between the sensing layer and the neural layer New neural neural layer 811 New neural network neural network contacts corresponding to each contact of the sensing layer 812 Machine learning unit 813 cerebral skin layer nerve called self-organization based on the probability scale connected between the dermal layer and the neural layer Database between Layers 814 New Neural Cerebral Layer 815 Cerebral Layer Contact 1101 Original Image Color 1102 Image Color After Scanning

Claims (7)

It is a method of constructing a machine learning model for regression analysis.
(1) To configure a machine learning system for regression analysis by installing software including at least regression analysis and self-organization based on the maximum probability scale on hardware including a processor and memory ;
(2) The above-mentioned machine learning system for regression analysis collaborates with hardware and software to create a linear regression line obtained from all the data in the processing area and the processing areas on both sides of the linear regression line. Perform self-organization processing based on the maximum probability scale only for the contained data to generate a new processing area;
(3) Regression analysis characterized in that the result of linear regression analysis is obtained by a process of repeating until the difference in data density from the previous iterative process in the processing area generated as described above becomes a predetermined value or less . How to configure a machine learning model. The probability scale of self-organization processing is normal distribution, multivariate normal distribution, logarithmic normal distribution, exponential distribution, t distribution, F distribution, X2 distribution, ^binomial distribution, negative binomial distribution, polynomial distribution, Poisson distribution, Maximum probability density consisting of probability distributions including at least one of Alan distribution, hypergeometric distribution, geometric distribution, traffic distribution, Weber distribution, triangular distribution, beta distribution, gamma distribution, probability distribution dispersion, mean square deviation, or As a measure of randomness including covariance
The machine learning model for regression analysis according to claim 1, wherein the maximum probability value of the self-organization is an average value or an expected value of a probability distribution obtained by the self-organization based on the scale of the maximum probability value. How to configure. The configuration of the machine learning model for regression analysis according to claim 1, wherein the self-organization based on the probability scale transfers the probability space in the maximum probability direction by processing the self-organization based on the probability scale. Method. The method for constructing a machine learning model for regression analysis according to claim 1, wherein the density value of the data in the processing region is set to the density of the data distribution with respect to the new processing region generated by the repeated processing. .. It is a component of machine learning for regression analysis.
(1) A machine learning system for regression analysis is configured by installing software including at least regression analysis and self-organization based on the maximum probability scale on hardware including a processor and memory.
(2) The above-mentioned machine learning system for regression analysis collaborates with hardware and software to obtain a linear regression line from all the data in the processing area and the processing areas on both sides of the linear regression line. Self-organization processing based on the maximum probability scale is performed only on the data contained in, and a new processing area is generated .
(3) Regression analysis characterized in that the result of linear regression analysis is obtained by a process of repeating until the difference in data density from the previous iterative process in the processing area generated as described above becomes a predetermined value or less . Machine learning component . A program for causing a processor to execute the method according to any one of claims 1 to 4. A general-purpose mobile terminal device equipped with the program according to claim 6.

Images