CN114548197A

CN114548197A - Clustering method based on self-discipline learning SDL model

Info

Publication number: CN114548197A
Application number: CN202011396635.XA
Authority: CN
Inventors: 顾泽苍
Original assignee: Individual
Current assignee: Individual
Priority date: 2020-11-26
Filing date: 2020-11-26
Publication date: 2022-05-27

Abstract

The invention relates to a clustering method based on an autonomous learning SDL model in the field of information processing, which is characterized by comprising the following steps: the clustering result of each class obtains the best clustering solution on the effects of function mapping and function Gaussian distribution according to the maximum probability scale of the probability space among the clustered feature vectors according to the scale of the distance of the probability space. The method is characterized in that: the SDL model clustering can consider the fusion of a function mapping model and a function Gaussian distribution model, can simulate a deep learning function mapping model, and can realize the high-precision image recognition capability and the high generalization capability of the function Gaussian distribution model. The method has no black box problem, does not need large hardware support, does not need marking of large data, and only needs training of small data, so the method has high performance, low import cost and convenient mass popularization.

Description

Clustering method based on autonomous learning SDL model

[ technical field ] A method for producing a semiconductor device

The invention belongs to a clustering method based on an autonomous learning SDL model in the field of artificial intelligence.

[ background of the invention ]

The "deep learning" (non-patent document 1) proposed by the Hinton team of the university of toronto, canada achieved excellent performance in the test data set of image classification of IMAGENET, attracting attention of the world, and therefore raised the climax of this artificial intelligence. Many researchers have been working on the control of autonomous vehicles using "deep learning" models. A typical technique is "driving at a day" (non-patent document 2).

Inventor Hinton as "deep learning" was declared on receiving Axios website interviews in 2017 at 9 months: "my view is to drop it (back-propagation) all away, re-start the stove", which is a dream break of the boltzmann machine by Hinton, the black box problem of "deep learning" is unsolvable and therefore not suitable for widespread popularization and eventually to end up.

Therefore, people need to find a new generation artificial intelligence model for replacing deep learning, and hopefully, a small-data, probabilistic and iterative machine learning model without the black box problem is obtained. However, the Capsule theory (non-patent document 3) by Hinton prediction does not produce the expected effect.

After deep learning is negated by the inventor, the algorithm school grows up, and one topic is: a new generation of artificial intelligence autonomous learning SDL model of a novel artificial intelligence neural network construction method (CN108510052A) is published.

[ non-patent document 1 ]

A.Krizhevsky，I.Sutskever，Geoffrey E.Hinton：“ImageNet Classification with Deep Convolutional Neural Networks”，Advances in Neural Information Processing Systems 25：pp1097-1105(2012).

[ non-patent document 2 ]

A.Kendall，J.Hawke：″Learning to Drive in Day″， arXiv：1807.00412v2.[cs.LG]11 Sep.(2018).

[ non-patent document 3 ]

S.Sabour，N.Frosst，Geoffrey E.Hinton：″Dynamic Routing Between Capsules″，arXiv：1710.09829v2.[cs.CV]7 Nov.(2017).

[ patent document 1 ]

(CN108510052A)

The deep learning model described above (non-patent document 1) requires exhaustive enumeration to obtain global optimum when solving a data set, which is an NPC problem in such a large combination space. Moreover, only the local optimal solution can be obtained by the SGD through the probability gradient descent method, and the global optimal solution is difficult to achieve. Moreover, the local optimal solution obtained by the SGD is random in application effect to deep learning, and it cannot be guaranteed that each SGD solution is the solution with the best application effect. The local optimal solution of SGD is very unstable because the global optimal solution is not available, and a distinct solution is obtained as long as the data fluctuates a little, which is the reason for the black box problem.

Moreover, huge hardware overhead is consumed in exhaustive calculation, the processing efficiency is extremely low, the hardware cost is very high, and the method is the reason for solving small tasks by a large model. In practical application, one deep learning algorithm personnel needs to be matched with 100 marking personnel, which is completely artificial intelligence, because the deep learning is a function mapping model, so that the application cost is very high. Further, the deep learning is limited by the application scene, and can be applied only to image recognition and voice recognition, and cannot be applied to industrial control, control of an autonomous vehicle, and the like.

The above (non-patent document 2) adopts a model-free depth reinforcement learning algorithm method, and uses Deep Deterministic Policy Gradients (DDPG) to solve the lane tracking task. In the face of complicated control of the automatic driving automobile, the method is easy to fall into the NPC problem of control and is difficult to be applied in practical engineering.

In the Capsule theory of the above (non-patent document 3), the weight value is increased by using the effective node information, and the weight value is decreased by using the ineffective node information, and as an iterative method, the result is calculated by a formulation method, and the like.

The above (patent document 1) introduces the most mathematically significant theory of the gaussian process, and is a probabilistic model, a small data model, and a highly iterative model. The application effect of an infinite data set corresponding to the function mapping can be obtained with only a small amount of data. The system scale can be infinitely expanded, and the computational complexity is close to linearity. Can be applied to any field. But does not have the feature of a deep learning function mapping model that can enlarge the interval between feature vectors.

[ summary of the invention ]

The first object of the present invention is: the deep learning of the mathematical calculation and simulation function mapping model is realized through the algorithm, so that the optimal solution of data training is not required to be obtained through combination, the processing efficiency of the system is improved, the hardware overhead is reduced, and the problem of black boxes is solved.

The second object of the present invention is: the method provides a method for combining the deep learning simulated by mathematical calculation with the SDL Gaussian distribution model and the probability distribution of the function with the mapping of the function, and has the advantages that the two characteristics can be exerted, the strongest artificial intelligence model is constructed at present, and the deep popularization of the artificial intelligence is promoted.

In order to achieve at least one of the above objects, the present invention proposes the following technical solutions.

A clustering method based on an autonomous learning SDL model at least has the following characteristics:

(1) the clustering feature vectors are subjected to a scale according to the probability space distance;

(2) clustering results of each class are according to the maximum probability scale of the probability space;

(3) the best solution of clustering is obtained on the effect of function mapping and function Gaussian distribution.

Moreover, the clustering algorithm of the SDL model is the fusion of a function mapping model and a Gaussian distribution model; performing optimal clustering on the feature vectors through probability scale self-organization and probability space distance; the clustering of the results of the individual probability spaces of the characteristic values is given directly; and a clustering algorithm for obtaining an optimal solution between the function mapping characteristic and the Gaussian distribution characteristic.

Moreover, the clustering algorithm of the SDL model is as follows: the mapping function comprises at least one of a linear function, a nonlinear function, a random function and a plurality of mixed mapping functions; or a clustering algorithm of the SDL model.

Moreover, the clustering algorithm of the SDL model is as follows: the mapping function is not only a classical linear function, a classical nonlinear function and a classical random function, and particularly, the mapping function is comprehensively constructed by considering the effect of deep learning on improving the accuracy of mode identification and combining a manual intervention method according to the characteristics of a solution solved by the SDG of the deep learning; the mapping function comprises components in a mathematical operation form, components with membership functions, regularly constructed components, at least one of clustering components of the SDL model, or a mixture of a plurality of components.

Moreover, the clustering algorithm of the SDL model is as follows: the probability space of the maximum probability is obtained by a probability scale self-organizing algorithm.

A clustering method based on an autonomous learning SDL model is characterized by comprising the following steps:

(1) performing probability scale self-organizing iteration on all data according to Euclidean distance between feature vectors to obtain two maximum probability Gaussian distributions

The maximum probability value and the maximum probability scale of (the maximum probability space);

(2) taking the two obtained maximum probability values as centers, taking the distance of the probability space as a scale with all the data which are not clustered, carrying out countermeasure in the two probability spaces, and taking the data within the two maximum probability scales as the final two clustering results;

(3) repeating the processes (1) to (2) until all the data are clustered.

Moreover, the clustering algorithm of the SDL model refers to: the mapping function is not only a classical linear function, a classical nonlinear function and a classical random function, and particularly, the mapping function is comprehensively constructed by considering the effect of deep learning on improving the accuracy of mode identification and combining a manual intervention method according to the characteristics of a solution solved by the SDG of the deep learning; the mapping function comprises components in a mathematical operation form, components with membership functions, regularly constructed components, at least one of clustering components of the SDL model, or a mixture of a plurality of components.

The invention provides a construction method based on algorithm simulation deep learning, which has the implementation effects that: the SDL model clustering can consider the fusion of a function mapping model and a function Gaussian distribution model, can simulate a deep learning function mapping model, and can realize the high-precision image recognition capability and the high generalization capability of the function Gaussian distribution model. The method has no black box problem, does not need large hardware support, does not need marking of large data, and only needs training of small data, so the method has high performance, low import cost and convenient mass popularization.

Drawings

FIG. 1 is a schematic diagram of a minimum-scale neural network configuration

FIG. 2 is an example of the relationship between the solution and the application effect of all SGDs obtained for an input message

FIG. 3 is a gray scale conversion image processing method

FIG. 4 is a diagram illustrating an image processing method for emphasizing frame information

FIG. 5 is another image processing method for emphasizing border information

FIG. 6 is a schematic diagram of a simulation deep learning based on SDL model

FIG. 7 is another structural diagram of simulation deep learning based on SDL model

FIG. 8 is a schematic diagram of various forms of mapping functions

FIG. 9 is a schematic of two overlapping Gaussian distributions

FIG. 10 shows training data of Image class for Image _ NET

FIG. 11 is a flow chart of the autonomous machine learning SDL clustering method

FIG. 12 is a schematic diagram of a depth separable convolution

FIG. 13 is a schematic diagram of a depth separable convolution when more features need to be extracted

Description of the symbols

I₁，I₂，I₃，I₄Is to input information

T₁，T₂，T₃，...，T₁₆Is a weight value

O₁，O₂，O₃，O₄Is the output information

(601) Is an input layer

(602) Is a Gaussian layer

(603) Is a function mapping layer

(604) Data set layer

(610) Image information

(611) Machine learning of SDL model composition

(612) Machine learning with SDL model construction

(613) Output data of Gaussian distribution

(701) Is an input layer

(702) Is a Gaussian layer

(703) Is a function mapping layer

(704) Data set layer

(710) Image information

(711) Machine learning with SDL model construction

(712) Machine learning with SDL model construction

(713) Output data of Gaussian distribution

G_ζA Gaussian distribution

G_ξAnother Gaussian distribution

Overlap of two gaussian distributions of ω

φ_maxζGaussian distribution G_ζMaximum value of

φ_maxξGaussian distribution G_ξMaximum value of

m_maxζGaussian distribution G_ζMaximum probability scale of

m_maxξGaussian distribution G_ξMaximum probability scale of

σ_ζGaussian distribution G_ζMaximum probability scale ofCompression value of

σ_ξGaussian distribution G_ξCompressed value of the maximum probability scale of

Initialization step

Two division steps

Data exchange procedure

Classification end judgment step

Probability space obtaining step

Detailed Description

The embodiments of the present invention will be described in further detail below with reference to the attached drawings, but the embodiments of the present invention are illustrative rather than restrictive.

First, some new definitions, new concepts and new formulas are introduced

[ probabilistic Scale self-organization ]

Setting a probability space:

[ EQUATION 1 ]

There is an initial region, the central value of which

And the variance of the Gaussian distribution calculated by the central value as the initial maximum severalScale of rate

To be provided with

As a central value, on a probability scale

The following iterations are performed for the reference:

[ equation 2 ]

The result of n iterations can obtain a maximum probability value close to the parent body of the probability space

Maximum probability scale

And a maximum probability space G⁽ⁿ⁾. Thereby constituting a probability scale self-organization.

[ migration of SDL model ]

The above-described SDK model can, by several iterations, necessarily migrate to converge within the region of greatest probability regardless of the initial region.

[ probabilistic space ]

Probability Space (Probability Space) as described herein: based on the theory that the "probability theory" of the Soviet Union mathematician Andrey Kolmogorov is based on the theory of the measure theory, the so-called probability space is a measurable space with the overall measure of "1". Theorem 1 can be generated according to this theory: "there is only one gaussian distribution in the probability space, so there is an infinite probability space in euclidean space.

[ probabilistic spatial distance ]

One point of euclidean space is measured to one probability space, or a measure between one probability space to another.

[ method for calculating a probabilistic spatial distance ]

Let the eigenvalues of the eigenvector V

Maximum probability value of probability space

And maximum probability scale

With another eigenvector W eigenvalue

Maximum probability value of probability space

Maximum probability scale

And feature vectors in Euclidean space

Characteristic value of

The distance G (V, W) between the euclidean space and the probability space can be unified as follows.

[ equation 3 ]

The following provides a method of opening a deep learning black box.

The well-known combination of more than 40 elements according to the theory of composition is the problem of NPC, which is not resolvable by turing machines. Based on this knowledge, we construct a minimum-scale neural network that can be exhaustively evaluated to obtain a globally optimal solution.

Fig. 1 is a schematic diagram of a minimum-scale neural network configuration.

As shown in fig. 1; i is₁，I₂，I₃，I₄Is input information, T₁，T₂，T₃，...，T₁₆Is a weight, i.e. the data set of the combined result, O₁，O₂，O₃，O₄Is the output information. According to the principle of neural network, then

[ EQUATION 4 ]

Order to

Then:

as shown in equation 4; this is a system of linear equations that should have a global optimal solution when the input information is equal to the output information. Therefore, the system is a stable system when the global optimal solution is obtained, and the black box problem does not exist.

A unique global optimal solution is found through an exhaustive method, the correctness of the formula 1 is proved, and meanwhile, according to the principle of a probability gradient descent method SGD, the solution of the SGD is solved through the exhaustive method. The number of local optimal solutions of the SGD, which is found to be a simple neural network, is random according to different input information. The small amount is hundreds, and the large amount can reach more than twenty thousand. The SGD solution may be advanced toward the globally optimal solution despite the occasional encounter of input information until the globally optimal solution is obtained. But this is extremely incidental. Mostly, the SGD method is extremely difficult to get across so many hills of local optimal solutions because the SGD solutions are too many. Therefore, the initiative of the SGD method to attempt to obtain the global best solution through SGD is not scientifically based and is a very wrong theory.

The purpose of opening a black box for deep learning is to try to break through the secret with good application effect of the deep learning in the fields of image recognition, image classification and the like. For each of various input information, we find the corresponding value of the application effect of the deep learning and all SGD local optimum solutions, and examine the relationship between the SGD solutions and the application effect.

Fig. 2 is an example of the relationship between the solutions of all SGDs obtained for a certain input information and the application effect.

As shown in fig. 2; from the first SGD solution obtained to the last 5,187 SGD solutions, the effect of the application is random and several times different. Therefore, the SGD method can not ensure that a global optimal solution can be obtained certainly, and can not ensure that the SGD solution is the solution with the best application effect of deep learning, so that the SGD method is a pseudo proposition.

After the black box of the neural network is opened, the mechanism of deep learning is thoroughly recognized through a large amount of data: the functional mapping implemented by the neural network through the combination of the neural networks for the input data may enlarge the interval of different feature vectors to hundreds or even thousands of times, or higher. The function mapping is random function mapping, and slight difference of input information can be mapped to results of distinct data sets through the random function mapping, so that the probability of misidentification of data of different types of data sets can be greatly reduced according to the Gaussian distribution theory. This is advantageous for image classification, as the accuracy of image recognition is improved. The prominent achievement of deep learning on the application effect is not determined by the form of the structure or weight generation of the neural network. But rather by the form of the functional mapping. The function mapping is mapped under independent single data, so that the result of correct mapping and the result of data matching can be obtained by the small difference between the feature vectors, which is the root of the deep learning to obtain the accuracy exceeding the traditional identification accuracy.

In order to improve the accuracy of image recognition, the method simulates deep learning, and in order to highlight the personalized features of the image, the image can be filtered through various templates or the recognized image is directly processed by an image processing algorithm. The processing method for highlighting the personalized features of the image is described below by taking a gray-scale value adjustment image processing method as an example, and the application of the method in the method for simulating deep learning based on the SDL model provided by the invention has novelty according to various methods disclosed in the past and belongs to the scope of the invention.

Fig. 3 is a gradation conversion image processing method.

Fig. 3(a) shows the original gray scale value of any 3 × 3 pixels in the original image. The maximum gray value among the original gray values of 3 × 3 pixels is exchanged with the central gray value as in fig. 3 (b). The minimum gray value among the original gray values of 3 × 3 pixels is swapped with the center gray value as in (c) of fig. 3. As shown in fig. 3 (d), the maximum probability value is calculated by using the SDL model for each gray value among the original gray values of 3 × 3 pixels, and the maximum probability value is exchanged with the center gray value.

In the method described above, the approximate probability value may be obtained by self-organizing the diagonal line of the original gray values of 3 × 3 pixels in fig. 3(a), and the pixel value of the center cross line with the maximum gray value and the minimum gray value among the original gray values of 3 × 3 pixels by using a probability scale.

Fig. 4 is an image processing method in which frame information is emphasized.

As shown in fig. 4 (a); the image is differentiated in the x direction and the y direction, and then the gray value of the original pixel is replaced by the result of multiplication according to the correspondence of each pixel by using the constants in the left 3 x3 grid and the right 3 x3 grid of the image 4 (a). Also, as shown in fig. 4 (b); the image is differentiated in the x direction and the y direction, and then the gray value of the original pixel is replaced by the result of multiplying the constants in the left 3 × 3 grid and the right 3 × 3 grid of the image 4 (b).

Fig. 5 is another image processing method of emphasizing the frame information.

As in fig. 4, the x-direction derivative result is multiplied by the template, as shown in fig. 5 (a), to obtain the processing effect of the vertical frame filter. As shown in fig. 5 (b), the y-direction derivative result is multiplied by the template, so that the processing effect of the horizontal frame filter can be obtained.

When the image is identified, one image can be converted into a plurality of images, and the Gaussian distribution of each characteristic vector can be formed, so that the identification rate is improved, and the quality of the image can be improved.

In particular, the number of feature vectors can be increased during image recognition, and the image recognition rate can be improved. Especially, the output of the convolutional neural network frequently used in deep learning can be directly input to the nodes of the input layer of the SDL model to be used as a group of new characteristic values, the number of characteristic vectors can be increased, the intervals among the characteristic vectors of different types of images can be increased, the scale of a data set can be increased, the classification precision of the images can be finally improved, and the precision of image identification can be improved.

The main convolution algorithm for deep learning is as follows:

1. gaussian convolution kernel

[ equation 5 ]

Corresponding to the pixels of the cells of each RGB color image, the processing results are accumulated and then averaged, and one pixel may be slid, two pixels may be slid, or three pixels may be slid, etc.

Roberts edge detection

[ equation 6 ]

Or

Prewitt edge detection

[ equation 7 ]

Or

Sobel edge detection

[ EQUATION 8 ]

Or

Scharr edge detection

[ equation 9 ]

Or

Laplacian operator

[ EQUATION 10 ]

Kirsch directional operator

[ equation 11 ]

The difference values in 8 directions are calculated, the maximum value is taken as the final output edge strength, and the corresponding direction is the edge direction.

8. Relief filter

[ EQUATION 12 ]

And filtering small-area noise in the image.

9. Edge reinforcement

[ equation 13 ]

10. Average filtering

[ equation 14 ]

11. Depth separable convolution

FIG. 12 is a schematic diagram of a depth separable convolution.

As shown in fig. 12; as with neural networks, deep separable convolution may also be used in the SDL model, where spatial convolution is performed while keeping the channels separated, followed by deep convolution. Taking an RGB image with an input image of 12 × 3 as an example, the normal convolution is to convolve 3 channels simultaneously with a convolution kernel. That is, 3 channels, after one convolution, output one number. While the depth separable convolution is divided into two steps: three convolutions are used to convolve the three channels, so that after one convolution 3 numbers are output. The three numbers output are passed through a 1x1x3 convolution kernel (poitwise kernel) to obtain a number. The depth separable convolution is actually achieved by two convolutions.

And step one, performing convolution on the three channels respectively and outputting the attributes of the three channels.

Secondly, convolving the three channels again by using a convolution kernel 1x1x3 to obtain a characteristic value data, wherein the characteristic value data at this time corresponds to 64 characteristic values, namely 8x8x 1.

FIG. 13 is a schematic diagram of a depth separable convolution when more features need to be extracted.

As shown in fig. 13; when more features need to be extracted, more 1x1x3 convolution kernels need to be designed (e.g., the 8x8x256 cube is drawn as 256 8x8x1, since they are not unity, representing 256 properties). In the SDL model, when more features need to be extracted, the processing method of the deep separable convolution is that the result output by the convolution neural network is directly input to each node corresponding to the input layer of the SDL model.

Deep learning is of interest to the world with excellent performance in Image classification of Image _ NET in 2012. In order to prove that the deep learning based on the algorithm simulation can achieve the same capacity of the common deep learning, a new-generation artificial intelligence model which is stronger is formed by taking Image classification of Image _ NET as an example, introducing an SDL function Gaussian distribution model and using a function mapping model for simulating the deep learning by the algorithm.

Fig. 6 is a schematic diagram of a simulation deep learning based on the SDL model.

As shown in fig. 6; (601) the input layer is mainly responsible for receiving image information (610) through an SDL (611) on each node of the input layer, and the input layer (602) is a Gaussian layer and is mainly responsible for obtaining the Gaussian distribution (613) of the characteristic values of the images through training of machine learning (612) on the same type of images input for multiple times. The function mapping layer (603) is mainly responsible for mapping the result (613) of the gaussian distribution obtained by the gaussian layer to the data set (604) by deep learning through algorithm simulation.

Here, the image information (610) may divide the image into η e δ pixel small regions, each of which finds a maximum fraction value of the region by the SDL model, and input the maximum fraction value to a corresponding node of the input layer. The maximum probability value of each small region constitutes a respective feature value, and the feature values of the maximum probability values of all the small regions of the image constitute a feature vector of the image.

In the SDL model, when more features need to be extracted, all small regions of an image are processed by a convolution algorithm as in deep learning, the processing result is input to corresponding nodes of an input layer, and for deep separation convolution, the output of a convolution neural network can be directly used as the input of the SDL model and directly input to corresponding nodes of the input layer.

In the following, we use mathematical formulas to express the principles of image classification and image recognition based on algorithm simulation deep learning.

The training data of alpha images formed by beta characteristic values input to each node of an input layer (601) is the same type of training set images obtained under different conditions, or different types of images mixed by data in different training sets, such as the same type of training images on Image _ NET, or different types of images are mixed together, and the intervals of different types of characteristic vectors are separated through training of characteristic vector classification of probability distance and maximum probability scale, hereinafter referred to as training images for short, and the expression is as follows:

[ equation 15 ]

Through the training of formula (15), each group of characteristic values

Gamma eigenvector groups consisting of beta maximum probability eigenvalues can be obtained from (equations 1-2):

[ equation 16 ]

Here: gamma is less than or equal to alpha, and the vector of the maximum probability scale can be obtained by (formula 1-2):

[ equation 17 ]

According to the definition of the probability space, each element is divided into

And m_maxi(i ═ 1, 2,. beta.) can be constructedA set of maximum probability spaces for the feature vectors:

[ equation 18 ]

Is composed of (formula 1 to 2)

And m_maxiIs to calculate the probability space s_maxijA constant of (1, 2.,. gamma.,. j.,. beta.) is calculated. There are homogeneous images and heterogeneous images in the gamma probability spaces, but the gaussian distribution intervals between the feature vectors of heterogeneous images must be separated. The difference between the deep learning and the novel SDL model for simulating the deep learning by the algorithm is that the deep learning only maps data into data set, the novel SDL model can separate the interval of Gaussian distribution of images of different classes, can map the Gaussian distribution into the data set, and has the characteristic of small data training effect on big data.

To improve the recognition accuracy, the maximum probability characteristic value of images of different classes is always desirable

And

the farther the probability space distance is, the better, this problem can be solved by functional mapping. Setting mapping function

The following inequalities may be satisfied:

[ equation 19 ]

Fig. 7 is a schematic diagram of another simulation deep learning based on the SDL model.

As shown in fig. 7; (701) the input layer is mainly responsible for receiving image characteristic information (710) through an SDL model (711) at each node of the input layer, the mapping function is 703, the mapping function is mainly responsible for mapping the characteristic information of the image output by the input layer to a data set layer (704), the Gaussian layer is 702, the Gaussian distribution of the maximum probability of the characteristic value of the image is obtained through training of machine learning (712) by mainly responsible for the data set of the same type of training image (formula 5) obtained by the data set layer (704).

When inputting images of different classes, the maximum probability Gaussian distribution (formula 18) of the characteristic values of the images of different classes is obtained through the training of machine learning (712), under the condition, if the Gaussian distributions of the images of two different classes are overlapped, the maximum probability scale values of the two Gaussian distributions are compressed, and finally, the maximum probability value which can represent the compressed Gaussian distribution (713) and the maximum probability scale values are obtained and are sent to each node of a Gaussian layer (702) to be used as output values.

Here, the image information (710) may divide the image into η e δ pixel small regions, and each small region obtains a maximum fraction value of the region by the SDL model and inputs the maximum fraction value to a corresponding node of the input layer. The maximum probability value of each small region constitutes a respective feature value, and the feature values of the maximum probability values of all the small regions of the image constitute a feature vector of the image.

The whole image can be directly subjected to feature extraction, and the result of the feature extraction is sent to the node of the input layer. Feature extraction can also be performed in each small region of the image through a convolution algorithm (formulas 5 to 14) commonly used in deep learning, and the extracted features of each small region are used as a group of feature values to form a new feature vector together with the feature values described above.

FIG. 8 is a schematic diagram of various forms of mapping functions.

May be a linear function, as shown in (a) of fig. 8; (801) for feature vectors consisting of individual feature valuesAnd (802) as a result of the mapping, the distance interval of the characteristic vector (801) can be arbitrarily increased by the mapping function, namely, the effect of amplifying the interval of the characteristic vector after the information is input and mapped to the data set layer by the mapping function imitating deep learning is achieved.

A nonlinear function is also possible, as shown in fig. 8 (b); (803) is a feature vector composed of individual feature values, and (804) is a mapped result, and the feature vector (803) can be mapped into a complex nonlinear result through a mapping function at will. The feature vectors mapped to the data set layer generate nonlinear effects by simulating the excitation function of deep learning, and are used for corresponding nonlinear data classification.

It may also be a random function, as shown in (c) of fig. 8; (805) for the eigenvectors formed from the individual eigenvalues, (806) for the mapped results, the eigenvectors (805) can be mapped arbitrarily by the mapping function into the results of a complex random function, which is a random permutation of the individual eigenvalues in the eigenvectors that mimics the random relationship between the SGD and the input information.

It is also possible to have a composite function of at least two of the three functions. As shown in (d) of fig. 8; (807) for the feature vector composed of each feature value, (808) for the mapped result, the feature vector (807) can be mapped into a complex function mapping result with the effect of random result and non-linear effect at will by the mapping function, which is also the characteristic of the function mapping result of deep learning.

Not only classical linearityThe function, the classical nonlinear function and the classical random function are combined with a manual intervention method to comprehensively construct a mapping function, and particularly, the mapping function is comprehensively constructed according to the characteristics of a solution solved by the SDG of the deep learning, and the effect of the deep learning on the improvement of the accuracy of the mode identification is considered. The mapping function has components in a mathematical operation form, membership function components, regularly constructed components and the like, and can meet the comprehensive function mapping model.

In combination with the mechanism of gaussian distribution, the gaussian distribution with the maximum probability of the eigenvalue derived from the training data of the object identified by the SDL model (612) or (712) output from each node of the gaussian layer (602) or (702), and all data within the maximum probability scale range of the maximum probability value in this gaussian distribution with the maximum probability should be mapped to the same data in the same data space according to the formula (3) of probability space distance.

The result of the gaussian distribution of the eigenvalues output from each node of the

gaussian layer

602 or 702 and obtained from the training data of the object identified by the

SDL model

612 or 712 may be regarded as infinite function mapping data and directly passed through the mapping function

And mapping the image to a data set layer, and on the judgment of the recognition result, solving the distance of a probability space according to the result of the amplified feature value Gaussian distribution, and taking the image corresponding to the data set with the closest distance as the recognition result.

Fig. 9 is a schematic illustration of two overlapping gaussian distributions.

As shown in fig. 9; two Gaussian distributions G obtained for two different classes of images_ζAnd G_ξThe overlapping part is omega, and the Gaussian distribution is shown as G_ζMaximum value of phi_maxζAnd the maximum probability scale value m_maxζAnd may represent a Gaussian distribution score G_ξMaximum value of phi_maxξMaximum probability scale value m_maxξ. The traditional method obtains the Gaussian distribution score G through probability scale self-organization_ζMaximum value of phi_maxζMaximum probability scale value m_maxζThen, the feature vector of the sample data (formula 18) and the Gaussian distribution G of the training data (formula 15) are calculated_ζThe minimum probability spatial distance of (2) can determine the recognition result of the image. In this case, the distance between the feature vectors of the images of different types is required to be as large as possible, which requires a technical effort in the quality of feature extraction or the number of feature values, and is limited in reality.

The functional mapping model does not need to take into account the maximization of the distance between the feature vectors of the different types of images, as long as the feature vector data of each mapped different type of image has to exist independently. Some processing is required for this purpose on the intervals of the feature vectors of different kinds of images.

As shown in fig. 9; the result of the gaussian distribution of the two different classes of images, which is the region of coincidence ω, is present, which means that there is a possibility of recognition errors, and if the data of the range of the largest probability scale is mapped onto the data set, there is a possibility of erroneous results of classifying the different classes of images into one class.

The invention proposes to put images of different classes together for training, and when the Gaussian distributions of the images of different classes shown in FIG. 9 coincide, the maximum probability scale value m of the two Gaussian distributions_maxζAnd m_maxξTo compress σ_ζAnd σ_ξValues of the range, m 'of the result of compressing the scale value of the maximum probability of the two Gaussian distributions can be obtained'_maxζAnd m'_maxξ. When the deep learning is simulated by the algorithm, the Gaussian distribution is used as mapping data, specifically, the maximum probability value of the Gaussian distribution and the compressed maximum probability scale m'_maxζAnd m'_maxξAs output of the mapping data, maximum probability scale m 'after compression of gaussian distribution'_maxζAnd m'_maxξAll data within (2) are to be regarded as being the same numerical value. I.e. in a gaussian distribution G_ζInterval of maximum probability of phi_maxζ-m’_maxζ≤ SP_ζ≤Φ_maxζ+m’_maxζSince it is the region of maximum probability, the same can be considered as the same by formula (3)The values, therefore, in this interval, can be treated as the same mapped value regardless of the value of the gaussian distribution. Also in the Gaussian distribution G_ξInterval of maximum probability of phi_maxξ-m’_maxξ≤SP_ξ≤Φ_maxξ+m’_maxξThe treatment method is the same as above.

If deep learning is completely simulated, the labeled data of big data is used for training, function mapping can be carried out without Gaussian distribution, and the data directly output from each node of the input layer 601 directly passes through the mapping function of the formula (10)

The feature vectors are mapped to a large data space (604).

In order to improve the generalization capability of machine learning, various kinds of characteristic vectors which are beneficial to expanding the distance between the characteristic vectors of the images in different classes can be used for distinguishing the images in different classes. Feature extraction of an image can be performed by the templates of fig. 3 to 5, or by extracting feature vectors of an image using convolution kernels (equations 5 to 14, fig. 12 and 13), or by a combination of a plurality of feature extraction methods. Finally, the processing result is input to the node of the input layer (601, 701) as a new feature value.

The following specifically describes the method for simulating deep learning by using an algorithm, which is provided by the present invention, for the problem of Image classification of Image _ NET.

Fig. 10 shows training data of Image _ NET for one Image class.

As shown in fig. 10; the training data is training data of goldfish images, and in order to achieve a higher-precision image classification effect, firstly, object images are dug out from the background through an artificial method, for example, the object images in fig. 9 are goldfishes, so that the goldfish images are dug out through the artificial method. This is also the process of telling the machine what is the object image by human intervention.

Next, the feature vector of the excavated image is obtained, and various feature vectors that can reflect the features of the object image and can distinguish other images can be generated by using the gray scale information of R, G, B, or a, B in Lab, or colors in other color spaces, the maximum probability value, the maximum probability scale, the maximum gray scale value, the minimum gray scale value, or the like, or the texture information obtained by deriving each color, or the like.

As shown in fig. 10, even though the target images are goldfish, different goldfish exist, and therefore, it is necessary to classify different goldfish into a probability space of the maximum probability by autonomous machine learning SDL clustering, and to mix data sets of other classes, and to deal with separation of intervals of feature vectors of different classes, so as to directly map gaussian distribution of each feature value to a data set or directly output the data set.

The algorithm for autonomous machine learning SDL clustering is as follows:

the traditional K-Means clustering is based on Euclidean distance as a scale, so that the probability space cannot be classified, the number of classes to be classified needs to be manually specified in advance, the best maximum probability self-discipline machine learning SDL clustering result cannot be obtained, and the Gaussian distribution of the maximum probability of the target function cannot be considered while the mapping of the target function is considered.

Fig. 11 is a flow chart of an autonomous machine learning SDL model clustering algorithm.

As shown in fig. 11; the clustering method is characterized in that data training does not need to be combined, the problem of black boxes does not exist, the optimal clustering result is obtained autonomously on the effects of function mapping and Gaussian distribution, the characteristic of function mapping of an objective function can be exerted on different types of feature vectors at small intervals, the identification result can be accurately obtained, and meanwhile, aiming at Image _ NET Image data in the type shown in figure 10, which is a single Image with large difference in color and texture, the clustering algorithm can obtain the optimal fusion result of a function mapping model and a Gaussian distribution model according to the extraction result of a given feature vector and given training data.

As shown in FIG. 11; the specific self-discipline machine learning SDL clustering steps are as follows:

STEP1 initialization STEP: respectively setting up databases which are not clustered yet

And already clustered databases

Initially putting all the feature vector data of the training data participating in the clustering

STEP2 probability scale self-organization STEP: for in database

All data are subjected to probability scale self-organizing iteration according to Euclidean distance between characteristic vectors to obtain Gaussian distribution capable of representing maximum probability

Constants of (maximum probability space), i.e. maximum probability values phi_maxζ(expected value), and maximum probability scale m_max(variance). Putting the data rejected in iteration back into the database

Once again for the database

Using self-organizing iteration using a probability scale to obtain another maximum probability value and a maximum probability scale (variance). Data culled in the iteration is also put back into the database

STEP3 produces two class STEPs: since the feature vector is high-dimensional data and only clustering based on euclidean space distance falls into a local optimum solution, the following processing is performed. By maximum probability value phi_maxIs a center, and a maximum probability scale m_maxTwo newly derived probability spaces represented, and a database

All data in (1) are calculated by using the distance of probability space, and countermeasures are carried out in the two probability spaces, and two maximum probability scales m are respectively used_maxThe data inside are stored in a database as the final two clustering results

In (1).

STEP4 probability scale modification STEP: aiming at two newly generated Gaussian probability spaces, the method also needs to be matched with a database

The probability space of the data of different training sets in (1) is processed by compressing the maximum probability scale of fig. 9, and a pair of compressed probability distribution data is stored in the database as the result of the function mapping data set

In this way, the function of high recognition accuracy of the function mapping can be exerted to the maximum, and the function of the maximum generalization capability of the gaussian distribution can be retained to the maximum.

STEP5 clustering completion judgment: and judging whether all vector data obtain clustering results, switching from Y to the next clustering ending STEP, and switching from N to STEP2 probability scale self-organization STEP.

STEP6 clustering completes the STEP.

In spite of the fact that the method shown in FIG. 6 is firstly performed on the characteristic values input to each node of the input layer, Gaussian distribution values are calculated through an SDL model, and then the maximum probability scale of probability spaces with overlapped Gaussian distributions of different classes is compressed and mapped to a data set; as shown in fig. 7, the characteristic values input to each node of the input layer are directly mapped to the data set layer, the gaussian distribution values at each node of the data set layer are calculated by the SDL model, then the maximum probability scale of the probability space with coincident gaussian distributions in different classes is compressed, and finally the maximum probability value and the maximum probability scale value are used as the output value of the SDL model.

The mapping mechanism of the target function of deep learning is focused on expanding the space of mapping data, namely, training values of large data obtained by combining complex neural networks can be correctly identified even if the distance between feature vectors of images of different classifications is small. Since each identified object is mapped into the data set, the generalization capability is poor, and all the states of the object image need to be labeled by big data to be practically applied.

The mechanism of gaussian distribution is to expand the distance between the feature vectors of different types of images as much as possible to improve the recognition accuracy of the images, and it is possible to have a very strong generalization capability by training of small data, but it is a problem that it is limited to expand the distance between the feature vectors of images as much as possible by improving the extraction quality of feature values. The Gaussian distribution has the characteristic of taking the inverse three aspects, and simultaneously, because the quality of the characteristic vectors can not ensure that the distance between the characteristic vectors of different types of images is large enough, different types of images or images in the same probability space can be caused, the interval between the characteristic vectors of different images is pulled open by finely compressing the maximum probability scale value of the Gaussian distribution, and the compressed Gaussian distribution result with the maximum probability is mapped to a large data space layer, namely an output layer. Therefore, the mapping effect of the target function can be obtained, and the method has the characteristic of small-data probability distribution of the target function.

The probability space of each maximum probability generated by the classification result is mapped by a mapping function in a function mapping layer (603)

The data is mapped into a data set (604).

In order to classify images with higher precision, the training data set and the test data set of the ImageNET can extract image outlines by considering the utilization of outline information of target images, and derivation is carried out from 8 directions, positions of derivative values of maximum density in different directions are connected according to the extracted outlines, and Gaussian distribution results of the direction and the length of the link are used as structural feature vectors to obtain more precise classification of the images so as to prevent the images from influencing the precision of image classification due to background noise.

Claims

1. A clustering method based on an autonomous learning SDL model at least has the following characteristics:

(3) the best solution of clustering is obtained on the effects of function mapping and function Gaussian distribution.

2. The clustering method based on the self-discipline learning SDL model as claimed in claim 1, wherein: the clustering algorithm of the SDL model is the fusion of a function mapping model and a Gaussian distribution model; performing optimal clustering on the feature vectors through probability scale self-organization and probability space distance; the clustering of the results of the individual probability spaces of the characteristic values is given directly; and a clustering algorithm for obtaining an optimal solution between the function mapping characteristic and the Gaussian distribution characteristic.

3. The clustering method based on the self-discipline learning SDL model as claimed in claim 1, wherein: the SDL model clustering algorithm is as follows: the mapping function comprises at least one of a linear function, a nonlinear function, a random function and a plurality of mixed mapping functions; or a clustering algorithm of the SDL model.

4. The clustering method based on the self-discipline learning SDL model as claimed in claim 1, wherein: the SDL model clustering algorithm is as follows: the mapping function is not only a classical linear function, a classical nonlinear function and a classical random function, and particularly, the mapping function is comprehensively constructed by considering the effect of deep learning on improving the accuracy of mode identification and combining a manual intervention method according to the characteristics of a solution solved by the SDG of the deep learning; the mapping function comprises components in a mathematical operation form, components with membership functions, regularly constructed components, at least one of clustering components of the SDL model, or a mixture of a plurality of components.

5. The clustering method based on the self-discipline learning SDL model as claimed in claim 1, wherein: the SDL model clustering algorithm refers to: the probability space of the maximum probability of the mapping function is obtained by a probability scale self-organizing algorithm.

6. A clustering method based on an autonomous learning SDL model is characterized by comprising the following steps:

(1) performing probability scale self-organizing iteration on all data according to Euclidean distances between the feature vectors to obtain the maximum probability value and the maximum probability scale of two maximum probability Gaussian distributions;

(3) repeating the processes (1) to (2) until all the data are clustered.

7. The clustering method based on the self-discipline learning SDL model as claimed in claim 6, wherein: the clustering algorithm of the SDL model is the fusion of a function mapping model and a Gaussian distribution model; performing optimal clustering on the feature vectors through probability scale self-organization and probability space distance; the clustering of the results of the individual probability spaces of the characteristic values is given directly; and a clustering algorithm for obtaining an optimal solution between the function mapping characteristic and the Gaussian distribution characteristic.

8. The clustering method based on the self-discipline learning SDL model as claimed in claim 6, wherein: the SDL model clustering algorithm is as follows: the mapping function may be a linear function, a nonlinear function, or a mixture of at least one or more random functions.

9. The clustering method based on the self-discipline learning SDL model as claimed in claim 6, wherein: the SDL model clustering algorithm is as follows: the mapping function is not only a classical linear function, a classical nonlinear function and a classical random function, and particularly, the mapping function is comprehensively constructed according to the characteristics of a solution solved by the deep learning SDG, the effect of the deep learning on improving the accuracy of the mode identification is considered, and the manual intervention technique is combined; the mapping function comprises at least one or more mixed components including components in mathematical operation form, components with membership functions and regularly constructed components.

10. The clustering method based on the self-discipline learning SDL model as claimed in claim 6, wherein: the SDL model clustering algorithm is as follows: the probability space of the maximum probability is obtained by autonomous machine learning SDL clustering.