Labeling method based on generative model
By using the generative Corr-LDA and Corr-HDP models, the sample likelihood probability is updated, solving the problem of incomplete labeling and achieving accuracy and completeness in labeling, which is suitable for the classification and retrieval of labeled data.
Patent Information
- Application Number
- CN202210482520.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-05
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-05-05
AI Technical Summary
In existing technologies, multi-label classification methods assume that the labels of the training data are completely appropriate, but in reality, the labels of the training data are often incomplete, which leads to the classification learning process being misled.
A generative model is adopted, and the Corr-HDP model is constructed through the Corr-LDA model. The likelihood probability of the samples is updated in a semi-supervised learning framework to enhance the correlation between labels and features, so as to achieve the accuracy and completeness of the annotation.
It effectively solves the problem of incomplete labeling, realizes label completion for incomplete labeled samples in the training set and accurate label prediction for new unlabeled samples, and improves the accuracy and completeness of labeling and classification.
Smart Images

Figure CN114997275B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of label data, and in particular to a label annotation method based on a generative model. Background Art
[0002] With the development of data acquisition, storage, and transmission technology, the amount of data has shown explosive growth in recent years. Effective data classification and retrieval methods can more conveniently manage large-scale and diverse data and discover useful information from them. It is worth noting that most existing data multi-classification methods, whether supervised learning or semi-supervised learning, assume that the labels of the training data are completely appropriate. This embodiment believes that such an assumption is difficult to achieve in reality, and even if it is not impossible to achieve, it is an unnecessary and unrealistic assumption. In reality, the labels of most training data are actually imperfect, and there are incomplete labels or noise labels. This embodiment mainly studies the incomplete labeling problem in multi-label classification, and attempts to solve this problem with a generative model. Data labeling is a time-consuming and laborious task. Due to human subjectivity, different focus points, lack of patience to label every detail, and even ignoring certain details due to human perception, even if you try hard to manually add labels to the data, some labels will still be missing. In this case, these labels are actually incomplete, and those who assume that the labels of the data in the training set are complete will obviously mislead the classification learning process.
[0003] Therefore, how to solve incomplete annotations in label classification has become a technical problem that needs to be solved urgently.
[0004] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention
[0005] The main purpose of the present invention is to provide a label annotation method based on a generative model, aiming to solve the technical problem of incomplete annotation that cannot be solved by the prior art.
[0006] To achieve the above object, the present invention provides a labeling method based on a generative model, the method comprising:
[0007] Obtaining a Corr-LDA model, and constructing the Corr-LDA model into a Corr-HDP model according to the broken stick theory;
[0008] In the Corr-HDP model, β is truncated to k dimensions;
[0009] A semi-supervised learning framework that obtains positive samples and unlabeled samples and uses the positive samples and unlabeled samples as training sets;
[0010] Applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework;
[0011] By updating the likelihood probabilities of all samples, the correlation between all labels and features is strengthened to achieve the accuracy and completeness of the labeled classification.
[0012] Optionally, the step of obtaining a Corr-LDA model and constructing the Corr-LDA model into a Corr-HDP model according to the broken stick theory includes:
[0013] Obtain the Corr-LDA model and use the notation of the broken stick theory to obtain the labels: β~Stick(α), π~DP(α^π, β);
[0014] Set the random variable y~Uniform(1,…,N i ) to associate features and labels in the same sample;
[0015] Given a sample i, generate N i Feature b i,l , Obey is a multinomial distribution with parameters;
[0016] For M i Any label among the labels, use y to randomly select N i One of the features generates the corresponding label w based on the hidden variable z that generates this feature. i,j ,in Obey is a multinomial distribution with parameters .
[0017] Optionally, the step of truncating β into k dimensions in the Corr-HDP model comprises:
[0018] In the Corr-HDP model, β is truncated into K dimensions, where π~DP(α π ,β) is approximately π~Dirichlet(α π β1,…,α π β K ).
[0019] Optionally, the step of obtaining positive samples and unlabeled samples and using the positive samples and unlabeled samples as a semi-supervised learning framework for training sets includes:
[0020] Get the incomplete labeled training set D, for each label L i ∈L, there exists a set pair P j×S j , where P j ∈D represents the positive sample set, S j ∈D represents a mixed sample set, w i,j =1→I i ∈P j , w i,j =0→I i ∈S j .
[0021] Optionally, before the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework, the step further includes:
[0022] In the training set, D can be represented as a set pair P for each j∈{1, 2, ..., M}. j ×S j , therefore, the following relationship is set:
[0023] I i ∈P j =>Pr[w i,j =1|b i ]=1
[0024] I i ∈P j ≠>Pr[w i,j =1|b i ]=0
[0025] in
[0026] Optionally, the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework includes:
[0027] Apply the semi-supervised learning framework to the Corr-HDP model and update each I k ∈S j The likelihood probability Pr[w k,j =1|b k ], and for each I i ∈P j (1≤j≤M) maintain Pr[w i,j =1|b i ]=1.
[0028] Optionally, after the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework, the method further includes:
[0029] When the updated value w of this label of a sample g,j In S j When all samples in the equation belong to the smallest t%, w g,j Set to zero.
[0030] In addition, to achieve the above purpose, the present invention also proposes a labeling device based on a generative model, the device comprising:
[0031] A model building module, used for obtaining a Corr-LDA model, and building the Corr-LDA model into a Corr-HDP model according to the broken stick theory;
[0032] A data truncation module, used for truncating β into k dimensions in the Corr-HDP model;
[0033] A sample acquisition module, used to acquire positive samples and unlabeled samples and use the positive samples and unlabeled samples as a semi-supervised learning framework for training sets;
[0034] A likelihood probability module, used to apply the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework;
[0035] The label annotation module is used to strengthen the correlation between all labels and features through the updated likelihood probabilities of all samples to achieve the accuracy and completeness of the label classification.
[0036] In addition, to achieve the above-mentioned purpose, the present invention also proposes a computer device, which includes: a memory, a processor, and a generative model-based labeling program stored in the memory and executable on the processor, wherein the generative model-based labeling program is configured to implement the generative model-based labeling method as described above.
[0037] In addition, to achieve the above-mentioned purpose, the present invention also proposes a medium, on which a label labeling program based on a generative model is stored, and when the label labeling program based on a generative model is executed by a processor, the steps of the label labeling method based on a generative model as described above are implemented.
[0038] The present invention obtains a Corr-LDA model and constructs the Corr-LDA model into a Corr-HDP model according to the broken stick theory; truncates β into k dimensions in the Corr-HDP model; obtains positive samples and unlabeled samples and uses the positive samples and unlabeled samples as a semi-supervised learning framework of a training set; applies the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework; and strengthens the correlation between all labels and features through the updated likelihood probabilities of all samples to achieve accuracy and completeness of labeling classification, which can not only complete the labels of incomplete labeled samples in the training set, but also accurately predict the labels of new completely unlabeled samples, thereby achieving the technical effect of solving incomplete labeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 It is a structural diagram of a label marking device based on a generative model in a hardware operating environment involved in an embodiment of the present invention;
[0040] Figure 2 It is a flowchart of the first embodiment of the labeling method based on the generative model of the present invention;
[0041] Figure 3 The graph models of Corr-LDA and Corr-HDP of the first embodiment of the labeling method based on the generative model of the present invention;
[0042] Figure 4 This is the graphical model of SSC-HDP of the first embodiment of the labeling method based on the generative model of the present invention. DETAILED DESCRIPTION
[0043] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0044] Reference Figure 1 , Figure 1 The present invention is a schematic diagram of a structure of a labeling device based on a generative model in a hardware operating environment according to an embodiment of the present invention.
[0045] like Figure 1As shown, the generative model-based labeling device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (Wireless-Fidelity, WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM) memory, or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.
[0046] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation on the generative model-based labeling device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.
[0047] like Figure 1 As shown, the memory 1005 as a storage medium may include an operating system, a data storage module, a network communication module, a user interface module, and a label marking program based on a generative model.
[0048] exist Figure 1 In the generative model-based labeling device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the generative model-based labeling device of the present invention can be set in the generative model-based labeling device, and the generative model-based labeling device calls the generative model-based labeling program stored in the memory 1005 through the processor 1001, and executes the generative model-based labeling method provided by the embodiment of the present invention.
[0049] The embodiment of the present invention provides a labeling method based on a generative model. Figure 2 , Figure 2 It is a flowchart of the first embodiment of the labeling method based on the generative model of the present invention.
[0050] In this embodiment, the label marking method based on the generative model includes the following steps:
[0051] Step S10: Obtain a Corr-LDA model, and construct the Corr-LDA model into a Corr-HDP model according to the broken stick theory.
[0052] It should be noted that the samples used in this embodiment are mainly images. In this embodiment, the sample set is represented as I, the feature dictionary of the sample is represented as V, and the label dictionary is represented as L. This embodiment uses the form of representing the features of each sample point and the annotations it contains. For each sample I i I i ∈I uses a one-dimensional vector b i b i =(b i,1 , b i,2 , …, b i,N )' as its characteristic representation, where b i,j Represents the jth feature V in the feature dictionary j ∈V in I i At the same time, use a vector w from M dimensions i =(w i,1 , w i,2 ,…,w i,m )' is used as the label of Ii, where w i,j ∈{0,1} indicates whether the jth labeled word in the label dictionary is in I i When I i When it is fully or partially marked, I i can be expressed as (b i , w i );I i When people are unlabeled samples, I i can be expressed as b i Therefore, in this embodiment, the incomplete labeled training set containing D samples is represented as At the same time, the unlabeled training set table containing U samples
[0053] Furthermore, the step of obtaining the Corr-LDA model and constructing the Corr-LDA model into a Corr-HDP model according to the broken stick theory includes: obtaining the Corr-LDA model, using the notation of the broken stick theory to obtain the label: β~Stick(α), π~DP(α^π, β); setting the random variable y~Uniform(1,…, N i ) to associate features and labels in the same sample; given a sample i, generate Ni feature Obey is a multinomial distribution with parameters; for M i Any label among the labels, use y to randomly select N i One of the features generates the corresponding label w based on the hidden variable z that generates this feature. i,j ,in Obey is a multinomial distribution with parameters .
[0054] It is understandable that the graph models of Corr-LDA and Corr-HDP are as follows Figure 3 As shown in the figure. The left side of the figure is the Corr-LDA model, and the right side of the figure is the Corr-HDP model. The difference between the two models is that in Corr-HDP, the prior β is obtained from a stick-breaking distribution, while in Corr-LDA, the prior is obtained from a finite Dirichlet distribution. The advantage of Corr-HDP is that it can dynamically and adaptively adjust the number of hidden variables according to the training set, rather than fixing the number of hidden variables as a priori.
[0055] In the specific implementation, it can be seen from the graphical model of Corr-HDP that the feature b and label w of the sample are simultaneously associated with the distribution variable z, which is obtained from a multinomial distribution with π as the parameter. The clustering mixed variable π is sampled from a global distribution G0 with β as the parameter. The stick-breaking theory is an effective method to establish the Dirichlet process. Using the notation of the stick-breaking theory, the following notation is obtained: β~Stick(α), π~DP(α^π, β). For the observed samples, we set a random variable y~Uniform(1,…, N i ) to associate features and labels in the same sample. Given a sample i, we first generate N i feature Obey is a multinomial distribution with parameters. Then, for M i Any label among the labels, use y to randomly select N i One of the features generates the corresponding label w based on the hidden variable z that generates this feature. i,j , Obey is a multinomial distribution with parameters. The following algorithm describes the generation process of the Corr-HDP model.
[0056]
[0057]
[0058] Step S20: truncating β into k dimensions in the Corr-HDP model.
[0059] Furthermore, the step of truncating β into k dimensions in the Corr-HDP model comprises: truncating β into K dimensions in the Corr-HDP model, where π~DP(α π ,β) is approximately π~Dirichlet(α π β1,…,α π β K ).
[0060] In practice, in order to make parameter estimation feasible, β needs to be truncated to K dimensions. Therefore, π~DP(α π ,β) can be considered as π~Dirichlet(α π β1,…,α π β K ). It should be noted here that truncating β to K dimensions in HDP is not equivalent to using K fixed topics in the Corr-LDA model. Using a truncated Dirichlet process (DP) as a prior can adaptively select the number of hidden variables based on the training set, and this number can be any value less than or equal to K. Using a Dirichlet distribution as a prior fixes the number of hidden variables to K.
[0061] Step S30: obtaining positive samples and unlabeled samples and using the positive samples and unlabeled samples as a semi-supervised learning framework of a training set.
[0062] Furthermore, the step of obtaining positive samples and unlabeled samples and using the positive samples and unlabeled samples as a semi-supervised learning framework for training sets includes: obtaining an incomplete labeled training set D, for each label L i ∈L, there exists a set pair P j ×S j , where P j ∈D represents the positive sample set, S j ∈D represents a mixed sample set, w i,j =1→I i ∈P j , w i,j =0→I i ∈S j .
[0063] Step S40: Apply the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework.
[0064] Furthermore, before the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework, it also includes: in the training set, D for each j∈{1, 2, ..., M} can be represented as a set pair P j ×S j , therefore, the following relationship is set:
[0065] I i ∈P j =>Pr[w i,j =1|b i ]=1
[0066] I i ∈P j ≠>Pr[w i,j =1|b i ]=0
[0067] in
[0068] Furthermore, the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework includes: applying the semi-supervised learning framework to the Corr-HDP model to update each I k ∈S j The likelihood probability Pr[w k,j =1|b k ], and for each I i ∈P j (1≤j≤M) maintain Pr[w i,j =1|b i ]=1.
[0069] Furthermore, after the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework, the method further includes: when the updated value w of the label of a certain sample g,j In S j When all samples in the equation belong to the smallest t%, w g,j Set to zero.
[0070] Step S50: By updating the likelihood probabilities of all samples, the correlation between all labels and features is strengthened to achieve the accuracy and completeness of the labeled classification.
[0071] In the specific implementation, the graph model of SSC-HDP is as follows Figure 4 Based on theoretical analysis, the main idea of semi-supervised learning here is to update each I under the structure of Corr-HDP. k ∈S j The likelihood probability Pr[w k,j =1|b k ], and for each I i ∈P j (1≤j≤M) maintain Pr[w i,j =1|b i ]=1. The following algorithm details the algorithm of the SSC-HDP model. By updating the likelihood probability of the label, the SSC-HDP model continuously strengthens the correlation between the label and the feature, and finally makes the label classification more accurate and complete. It can be seen that the information of the unlabeled data can be seamlessly added to the SSC-HDP algorithm. After the initial Corr-HDP model is established, the parameters of the model are used to calculate the likelihood probability of each label of each sample in the unlabeled data set U, and all these unlabeled samples are merged into the mixed sample set S j (1≤j≤M). The information of unlabeled data can be utilized by mixing sample sets to update the parameters of Corr-HDP in continuous iterations.
[0072]
[0073]
[0074] This embodiment obtains a Corr-LDA model and constructs the Corr-LDA model into a Corr-HDP model according to the broken stick theory; truncates β into k dimensions in the Corr-HDP model; obtains positive samples and unlabeled samples and uses the positive samples and unlabeled samples as a semi-supervised learning framework of the training set; applies the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework; through the updated likelihood probability of all samples, strengthens the correlation between all labels and features to achieve accuracy and completeness of labeling classification, can not only complete the labels of incomplete labeled samples in the training set, but also accurately predict the labels of new completely unlabeled samples, thereby achieving the technical effect of solving incomplete labeling.
[0075] In addition, an embodiment of the present invention further proposes a medium on which a label marking program based on a generative model is stored. When the label marking program based on a generative model is executed by a processor, the steps of the label marking method based on a generative model as described above are implemented.
[0076] The embodiments or specific implementation methods of the label marking device based on the generative model of the present invention can refer to the above-mentioned method embodiments, which will not be repeated here.
[0077] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0078] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.
[0079] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.
[0080] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A labeling method based on a generative model, characterized in that: The method comprises: Obtaining a Corr-LDA model, and constructing the Corr-LDA model into a Corr-HDP model according to the broken stick theory; In the Corr-HDP model, β is truncated to k dimensions; A semi-supervised learning framework that obtains positive samples and unlabeled samples and uses the positive samples and unlabeled samples as training sets; Applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework; By updating the likelihood probabilities of all samples, the correlation between all labels and features is strengthened to achieve the accuracy and completeness of the labeled classification, where the samples are images; The step of obtaining the Corr-LDA model and constructing the Corr-LDA model into a Corr-HDP model according to the broken stick theory includes: Obtain the Corr-LDA model and use the notation of the broken stick theory to obtain the labels: β~Stick(α), π~DP(α^π, β); Set the random variable y~Uniform(1,…,N i ) to associate features and labels in the same sample; Given a sample i, generate N i Feature b i,l , Obey is a multinomial distribution with parameters; For M i Any label among the labels, use y to randomly select N i One of the features generates the corresponding label w based on the hidden variable z that generates this feature. i,j ,in Obey is a multinomial distribution with parameters; The step of truncating β into k dimensions in the Corr-HDP model comprises: π In the Corr-HDP model, β is truncated into K dimensions, where π~DP(α, β) is approximately ππ Like π~Dirichlet(αβ1,…,αβ K ); The step of obtaining positive samples and unlabeled samples and using the positive samples and unlabeled samples as a semi-supervised learning framework for training sets includes: Get the incomplete labeled training set D, for each label L i ∈L, there exists a set pair P j ×S j , where P j ∈D represents the positive sample set, S j ∈D represents a mixed sample set, 2. The method according to claim 1, characterized in that Before the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework, the method further includes: In the training set, D can be represented as a set pair P for each j∈{1, 2, ..., M}. j ×S j , therefore, the following relationship is set: I i ∈P j =>Pr[w i,j =1|b i ]=1 I i ∈P j ≠>Pr[w i,j =1|b i ]=0 in 3. The method according to claim 1 or 2, characterized in that The step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework includes: Apply the semi-supervised learning framework to the Corr-HDP model and update each I k ∈S j The likelihood probability Pr[w k,j =1|b k ], and for each I i ∈P j (1≤j≤M) maintain Pr[w i,j =1|b i ]=1.
4. The method according to claim 3, characterized in that After the step of applying the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework, the method further includes: When the updated value w of this label of a sample g,j In S j When all samples in the t% are the smallest, w g,j Set to zero.
5. A labeling device based on a generative model, characterized in that: The device comprises: A model building module, used for obtaining a Corr-LDA model, and building the Corr-LDA model into a Corr-HDP model according to the broken stick theory; A data truncation module, used for truncating β into k dimensions in the Corr-HDP model; A sample acquisition module, used to acquire positive samples and unlabeled samples and use the positive samples and unlabeled samples as a semi-supervised learning framework for training sets; A likelihood probability module, used to apply the semi-supervised learning framework to the Corr-HDP model to update the likelihood probability of each sample in the semi-supervised learning framework; The label annotation module is used to strengthen the correlation between all labels and features through the updated likelihood probabilities of all samples to achieve the accuracy and completeness of the label classification, where the samples are images; The model building module is also used to obtain the Corr-LDA model, using the notation of the broken stick theory to obtain the label: β~Stick(α), π~DP(α^π, β); set the random variable y~Uniform(1,…, N i ) to associate features and labels in the same sample; given a sample i, generate N i Feature b i,l , Obey is a multinomial distribution with parameters; for M i Any label among the labels, use y to randomly select N i One of the features generates the corresponding label w based on the hidden variable z that generates this feature. i,j ,in Obey is a multinomial distribution of parameters; a data truncation module is also used to truncate β into K dimensions in the Corr-HDP model, where π~DP(α π ,β) is approximately π~Dirichlet(α π β1,…,α π β K ); The sample acquisition module is also used to obtain the incomplete labeled training set D. For each label L i ∈L, there exists a set pair P j ×S j , where P j ∈D represents the positive sample set, S j ∈D represents a mixed sample set, w i,j =1→I i ∈P j , w i,j =0→I i ∈S j .
6. A labeling device based on a generative model, characterized in that: The device includes: a memory, a processor, and a generative model-based labeling program stored in the memory and executable on the processor, wherein the generative model-based labeling program is configured to implement the steps of the generative model-based labeling method as described in any one of claims 1 to 4.
7. A medium, characterized in that The medium stores a label marking program based on a generative model, and when the label marking program based on a generative model is executed by a processor, the steps of the label marking method based on a generative model as described in any one of claims 1 to 4 are implemented.
