Generation method for ai learning material set

Abstract

The present application provides a generation method for an AI learning material set. The method comprises: acquiring an original learning material set, an algorithm and a factor relationship table in a target application scenario; performing factor item labeling on each piece of learning material in the original learning material set on the basis of the algorithm, and on the basis of the factor relationship table, determining the degree of association between a target factor item needing to be learned and other factor items in a model training process adapted to the target application scenario; and screening the original learning material set on the basis of the determined degree of association to obtain a learning material set obtained after screening, wherein the learning material set obtained after screening is used as input data for model training in the target application scenario, so as to generate a trained model. According to the present application, an automation characteristic is achieved on the basis of factor relationship table derivation, and additionally, material sets are minimized while high accuracy is achieved, such that the material sets are more organized and satisfy the training requirements (such as terminal training requirements) in a target application scenario.

Description

A method for generating AI learning data sets

[0001] This application claims priority to Chinese patent application number CN 202410309250.7, filed on December 20, 2023, entitled “A Method for Generating an AI Learning Data Set”, the disclosure of which is incorporated herein by reference. Technical Field

[0002] The present application relates to the field of artificial intelligence technology, and more specifically, to a method for generating an AI learning data set. Background Art

[0003] In recent years, with the widespread adoption and rapid development of artificial intelligence (AI), various AI-integrated applications have also experienced rapid growth. Traditionally, the use of AI requires users to first train a model. This trained model is then deployed in an inference system, which then implements the trained model into specific AI applications.

[0004] Currently, AI learning datasets mainly emphasize non-prior knowledge of large texts (big data). There are generally no standards for the establishment, screening, and processing of the entire dataset, which makes it difficult for the subsequently trained models to meet the needs of actual applications. Summary of the Invention

[0005] The embodiments of the present application at least provide a method for generating an AI learning data set, which determines a learning data set that better meets the AI ​​training requirements by the degree of correlation between factor items, and is more practical and universal.

[0006] In a first aspect, an embodiment of the present application provides a method for generating an AI learning data set, comprising:

[0007] Obtaining an original learning data set, an algorithm, and a factor relationship table for a target application scenario; the factor relationship table includes multiple factor items adapted to the target application scenario and correlation relationships between the multiple factor items;

[0008] Labeling each learning material in the original learning material set with factor items based on the algorithm, and determining, based on the factor relationship table, the degree of correlation between the target factor item to be learned and other factor items in the model training process applicable to the target application scenario;

[0009] The original learning data set is screened based on the determined degree of association to obtain a screened learning data set, and the screened learning data set is used as input data for model training in the target application scenario to generate a trained model.

[0010] Optionally, screening the original learning data set based on the determined association degree includes:

[0011] Determining, based on the degree of association, a minimum amount of learning materials required to learn the target factor item during the model training process of the target application scenario;

[0012] The original learning material set is screened according to the minimum number of learning materials to obtain a screened learning material set.

[0013] Optionally, the minimum amount of learning materials required to learn the target factor item is determined according to the following steps:

[0014] In response to the correlation degree belonging to a target correlation threshold range in a preset multi-level correlation threshold range, determining a statistical analysis amount within the target correlation threshold range;

[0015] The minimum amount of learning materials is determined based on the statistical analysis amount.

[0016] Optionally, screening the original learning material set according to the minimum amount of learning materials includes:

[0017] Based on the factor items marked on each learning material in the original learning material set, target learning materials marked with target factor items are selected from the original learning material set; the number of target learning materials is the minimum number of learning materials required to learn the target factor items;

[0018] The filtered learning material set is determined based on the selected target learning material.

[0019] Optionally, determining the statistical analysis amount within the target association threshold range includes:

[0020] Obtain the test conditions required for preliminary test analysis;

[0021] Determining whether each learning material within the target association threshold range passes the verification condition based on a statistical verification sample value table lookup method or an algorithm verification method;

[0022] When it is determined that the test condition is met, the statistical analysis amount is determined.

[0023] Optionally, determining the statistical analysis amount includes:

[0024] Obtaining a statistical hypothesis error rate and a priori sample characteristics preset for the original learning data set;

[0025] The statistical analysis quantity is determined based on the statistical hypothesis error rate and the prior sample characteristics.

[0026] Optionally, screening the original learning material set according to the minimum amount of learning materials includes:

[0027] The factor items in the factor relationship table that are covariant or correlated with the target factor item are used as covariant factor items, and the other factor items are used as random factor items;

[0028] stratifying the original learning data set according to the random factor item to obtain a stratified learning data set;

[0029] Select learning materials whose total amount is consistent with the minimum amount of learning materials from the stratified learning material set.

[0030] Optionally, the method further includes:

[0031] Obtaining a focused viewpoint image captured in a scenario applicable to the trained model;

[0032] The filtered learning data set is expanded or replaced based on the focused viewpoint image to obtain an updated learning data set.

[0033] Optionally, the method further includes:

[0034] Obtaining a learning rule to be added; the learning rule at least includes the degree of correlation between the characteristics of the factor item to be added and other factor items;

[0035] The filtered learning data set is updated based on the learning rule to obtain an updated learning data set.

[0036] Optionally, the method further includes:

[0037] Obtaining physical parameters of the terminal device to which the trained model is applicable;

[0038] The filtered learning data set is updated based on the physical parameters to obtain an updated learning data set adapted to the terminal device.

[0039] In a second aspect, the present application also provides a model training method, comprising:

[0040] Obtaining a filtered learning data set generated based on the method for generating an AI learning data set according to any one of the first aspect and various embodiments thereof;

[0041] At least one round of model training is performed based on the screened learning data set to obtain a trained model.

[0042] In a third aspect, the present application further provides a device for generating an AI learning data set, comprising:

[0043] An acquisition module, configured to acquire an original learning data set, an algorithm, and a factor relationship table for a target application scenario; the factor relationship table includes multiple factor items adapted to the target application scenario and the relationships between the multiple factor items;

[0044] a determination module for labeling factor items of each learning material in the original learning material set based on the algorithm, and determining, based on the factor relationship table, the degree of correlation between the target factor item to be learned and other factor items in the model training process applicable to the target application scenario;

[0045] A screening module is used to screen the original learning data set based on the determined degree of association to obtain a screened learning data set, and the screened learning data set is used as input data for model training in the target application scenario to generate a trained model.

[0046] In a fourth aspect, the present application also provides a model training device, comprising:

[0047] An acquisition module, configured to acquire a filtered learning data set generated based on the method for generating an AI learning data set according to any one of the first aspect and various embodiments thereof;

[0048] The training module is used to perform at least one round of model training based on the screened learning data set to obtain a trained model.

[0049] In the fifth aspect, the present application also provides an electronic device comprising: a processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the method for generating an AI learning data set as described in any one of the first aspect and its various embodiments and the method for model training described in the second aspect are executed.

[0050] In the sixth aspect, the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it executes the method for generating an AI learning data set as described in any one of the first aspect and its various embodiments and the model training method described in the second aspect.

[0051] The above-mentioned AI learning data set generation method and device, model training method and device are used. When the generation method obtains the original learning data set, calculation table and factor relationship table under the target application scenario, the factor items of each learning data in the original learning data set can be labeled based on the algorithm, and the degree of correlation between the target factor item and other factor items required for learning in the model training process applicable to the target application scenario is determined based on the factor relationship table, and then the original learning data set is screened based on the degree of correlation to obtain a screened learning data set. The screened learning data set obtained here is based on the degree of correlation between the key factor items (such as the target factor items) and other factor items determined by the model training, and the data is screened. Since this method is based on the factor relationship table and thus has an automated feature, it will obtain the smallest possible data set with higher accuracy. This makes the screened learning data set more orderly and adaptable to the training requirements of the target application scenario (such as terminal training requirements), and has better practicality and versatility.

[0052] Other advantages of the present application will be explained in more detail with reference to the following description and accompanying drawings.

[0053] It should be understood that the above description is only an overview of the technical solution of this application, so that one can have a general understanding of the technical means of this application and then implement it in accordance with the contents of the description. In order to make the above and other purposes, features and advantages of this application more obvious and easy to understand, the following examples are used to illustrate the specific implementation methods of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of this specification. These drawings illustrate embodiments that comply with the present application and are used together with the specification to illustrate the technical solutions of the present application. It should be understood that the drawings only illustrate certain embodiments of the present application and should not be regarded as limiting the scope of protection. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings. In the drawings:

[0055] FIG1 shows a flow chart of a method for generating an AI learning data set provided in an embodiment of the present application;

[0056] FIG2 shows a flow chart of a model training method provided in an embodiment of the present application;

[0057] FIG3 shows a schematic diagram of a device for generating an AI learning data set provided in an embodiment of the present application;

[0058] FIG4 shows a schematic diagram of a model training device provided in an embodiment of the present application;

[0059] FIG5 shows a schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] The following describes exemplary embodiments of the present application in more detail with reference to the accompanying drawings. Although the accompanying drawings illustrate exemplary embodiments of the present application, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0061] In the description of the embodiments of the present application, it should be understood that terms such as "including" or "having" are intended to indicate the presence of disclosed features, numbers, steps, actions, components, parts, or a combination thereof in the present specification, and do not exclude the possibility of the presence of one or more other features, numbers, steps, actions, components, parts, or a combination thereof.

[0062] Unless otherwise specified, “ / ” means or. For example, A / B can mean A or B. “And / or” in this article is only a way to describe the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0063] The terms "first," "second," etc., are used solely to distinguish identical or similar technical features for ease of description and should not be construed as indicating or implying the relative importance or quantity of these technical features. Thus, a feature defined by "first," "second," etc., may explicitly or implicitly include one or more of such features. In the description of the embodiments of this application, unless otherwise specified, the term "plurality" means two or more than two.

[0064] Research has found that current AI learning data sets mainly emphasize non-prior knowledge of large texts (big data). There are too many data sets, and there are generally no standards for the establishment, screening, and processing of the entire data set. This makes it difficult for the subsequently trained models to meet the needs of actual applications.

[0065] Especially for terminal applications, existing technical models are too large overall, with too many data sets, and various problems will arise when relying on a remote large model host for training. At this time, these problems can be improved by pruning the large model and applying it to the terminal. However, current methods such as pruning still result in the model being too large or the algorithm being complex after pruning, or the terminal requirements are too high, unable to load, and the calculation response time is too long, which easily leads to problems in application.

[0066] In order to at least partially solve one or more of the above-mentioned problems and other potential problems, the present application provides a method and device for generating an AI learning data set, and a method and device for model training, so as to determine a learning data set that better meets the AI ​​training needs through the degree of correlation between factor items, with the goal of reducing and shrinking the data set so that the trained model as a whole can be more convenient for the terminal to use.

[0067] To facilitate understanding of this embodiment, we first provide a detailed introduction to a method for generating an AI learning dataset disclosed in an embodiment of this application. The method for generating an AI learning dataset provided in this embodiment of this application is generally performed by an electronic device with certain computing capabilities, such as a server or other processing device. In some possible implementations, this method for generating an AI learning dataset can be implemented by a processor invoking computer-readable instructions stored in a memory.

[0068] 1 , which shows a flow chart of a method for generating an AI learning data set provided in an embodiment of the present application, the method includes the following steps S101 to S103:

[0069] S101: Obtaining an original learning data set, an algorithm, and a factor relationship table for a target application scenario; the factor relationship table includes multiple factor items adapted to the target application scenario and the relationships between the multiple factor items;

[0070] S102: labeling each learning material in the original learning material set with factor items based on an algorithm, and determining, based on a factor relationship table, the degree of correlation between the target factor item to be learned and other factor items in a model training process suitable for the target application scenario;

[0071] S103: The original learning data set is screened based on the determined degree of association to obtain a screened learning data set, which is used as input data for model training in a target application scenario to generate a trained model.

[0072] In order to facilitate understanding of the generation method of the AI ​​learning data set provided in the embodiment of the present application, the application scenario of the method is briefly introduced first. The generation method provided in the embodiment of the present application can be mainly used in the field of artificial intelligence technology, for example, it can be applied to various artificial intelligence sub-fields such as image recognition and face detection, and no specific restrictions are made here. In particular, before the need to deploy the relevant neural network model on the terminal, it is often possible to obtain a learning data set that matches the applicable target application environment and has been effectively screened through the generation method provided in the embodiment of the present application, so that the model obtained by subsequent training can be better applied on the terminal or other small computing power devices.

[0073] Based on different target application scenarios, the original learning data sets obtained here are also different. For example, for the application scenario of high-speed mobile safe driving, the original learning data set can be relevant road images or other relevant detection data (such as sound or other time series data that can be collected by sensors). There are no specific restrictions on the form of the learning data and its combination method. The following examples are mainly based on image data.

[0074] In practical applications, as long as the quality standards meet the requirements, they can be used as original learning materials. For example, for driving needs, store camera footage on various routes, driving images of other vehicles, etc., or storage databases of these data, or storage databases of processed data, as long as they meet the preset quality standards, they can be data providers, and there is no specific limitation here.

[0075] The algorithms include but are not limited to sampling algorithms. In actual terminal applications, the algorithms here can be determined by the terminal based on a certain application requirement. For example, when face recognition is required in a rainy driving environment, it can be determined which algorithm is needed to extract the algorithm method required for face recognition from the image. Based on the corresponding algorithm, each learning data in the original learning data set can be labeled with factor items, that is, for example, it can be labeled to affect safety.

[0076] The factor relationship table can be extracted from a large factor relationship table database, mainly including a variety of factor items adapted to the target application scenario and the correlation between the multiple factor items. Here we still take the target application scenario of high-speed mobile safe driving as an example. The factor items may include, for example, "affecting images", "affecting safety", "humidity", "color temperature" and other image recognition influencing factor items. The "hope to learn factor item" is an identifiable driving risk factor item, which can be used as a target factor item here. The relationship between the relevant factor items can be, for example, a correlation relationship, or a logical relationship (functional relationship, causal relationship, covering relationship, indirect relationship, parallel relationship, etc.).

[0077] The factor relationship table can be used to determine the degree of correlation between the target factor item and other factor items to be learned during model training for the target application scenario. The following is a specific example of determining the degree of correlation between the target factor item and other factor items based on the factor relationship table.

[0078] After determining the degree of association, the original learning data set can be filtered based on the degree of association to obtain a filtered learning data set. The higher the degree of association, the more learning data that needs to be extracted to a certain extent, and conversely, the lower the degree of association, the less learning data that needs to be extracted to a certain extent. However, the specific amount of extraction that is sufficient to meet the needs of training the model in the target application scenario will be a problem to be solved in the embodiments of this application.

[0079] In the embodiment of the present application, the original learning data set can be screened according to the following steps:

[0080] Step 1: Determine the minimum amount of learning data required to learn the target factor item during the model training process of the target application scenario based on the degree of association;

[0081] Step 2: Filter the original learning data set according to the minimum number of learning materials to obtain a filtered learning data set.

[0082] Here, the minimum amount of learning materials required for the learning objective factor items can be determined based on the degree of association, and then the original learning material set can be screened according to the minimum amount of learning materials.

[0083] The minimum amount of learning materials required for learning objective factor items is determined as follows:

[0084] Step 1: In response to the correlation degree falling within a target correlation threshold range in a preset multi-level correlation threshold range, determining a statistical analysis amount within the target correlation threshold range;

[0085] Step 2: Determine the minimum amount of learning materials based on statistical analysis.

[0086] After determining the degree of correlation between other factor items and the target factor item, the statistical analysis amount under each correlation threshold range can be determined according to different correlation degrees. For example, there is a factor item A that is highly correlated with the target factor item, and there is a factor item B that is moderately correlated with the target factor item. Here, the first statistical analysis amount required under the first correlation threshold range indicated by the high correlation can be determined, and the second statistical analysis amount required under the second correlation threshold range indicated by the moderate correlation can also be determined. Then, these statistical analysis amounts are combined to obtain the minimum amount of learning materials required to learn the target factor item.

[0087] When actually determining the statistical analysis amount under each correlation threshold range, the following steps may be used to determine the amount, wherein the target correlation threshold range is used as a specific example of each correlation threshold range:

[0088] Step 1: Obtain the test conditions required for preliminary test analysis;

[0089] Step 2: Determine whether each learning material within the target correlation threshold has passed the verification condition based on a statistical verification sample value table lookup method or an algorithm verification method;

[0090] Step 3: After the test conditions are determined to be passed, determine the statistical analysis quantity.

[0091] The reason why the embodiment of the present application adopts the verification and judgment method to determine the statistical analysis quantity is mainly because the large model training process is generally an iteration of statistical methods such as cluster analysis, factor analysis, convolution, and fitting.

[0092] According to the central limit theorem, the law of large numbers, etc., the above factors will produce different types of distribution characteristics due to the different clustering analysis algorithms adopted, and therefore will have statistical characteristics (such as no-parent number test, Bayesian estimation, etc.).

[0093] Furthermore, the commonly used underlying fitting algorithms for segmented AI, as documented in current literature and online materials, include regression, regularization algorithms, ensemble algorithms, decision tree algorithms, support vector machines, dimensionality reduction algorithms, clustering algorithms, Bayesian algorithms, artificial neural networks, deep learning, linear regression, logistic regression, decision trees, naive Bayes, support vector machines (SVM), ensemble learning, K-nearest neighbor algorithms, K-means algorithms, reinforcement learning networks (Deep Q-Networks, DQN), ridge regression, Lasso regression, decision tree regression, random forest regression, gradient boosting regression, support vector regression, XGBoost regression, LightGBM regression, neural network regression, etc. If any of the above algorithms is used to fit a specified factor relationship when calibrating a large model dataset, it means that the corresponding algorithm relationship is significant and has passed the test (if there is no preset test value, there will also be a minimum fit among all similar algorithms, etc.).

[0094] Therefore, when the sample after sampling using the embodiment of the present application (which complies with the sampling composition provisions of the present application) cannot fit the specified factor relationship, it means that this sample has statistical characteristics consistent with the original large model data set, but the number of samples may be too small to obtain the specified relationship significantly through the same algorithm, and the test passes (or reaches the lowest fit of all similar regressions).

[0095] At this time, the factor statistical characteristics of the large model can be obtained from the original prior distribution of the large model, and the minimum sample number estimate (corresponding to the minimum amount of learning data) can be obtained by reverse deduction from the corresponding algorithm or verification formula, which can then be applied to the learning data set used for large model training that meets the above assumptions.

[0096] The minimum sample size estimate here must meet two conditions, namely, avoiding the following two types of errors: Type I error: when the null hypothesis H0 matches the actual situation, the test results reject it, known as the "rejection of the true" error. Type II error: when the null hypothesis H0 does not match the actual situation, the test results cannot reject it, known as the "taking the false" error. Type I error is primarily due to inferences from fitting functions and algorithms (i.e., reaching the minimum number of functional or statistically significant values), while Type II error is primarily due to probability statistics.

[0097] In statistics, various test methods are used to determine whether the above errors are likely to occur. An error rate is first assumed (reflected in the large model as the significance of the built-in corresponding algorithm relationship, the test passing standard, or the minimum fit of all similar algorithms, etc.), and then the a priori sample characteristics (the statistical characteristics of the factor in the large model), the proposed algorithm (including the test method) and the data screening method are used to reversely deduce the minimum sample size using the algorithm or test function, and then use it to collect data or conduct experiments. When actually reversing the minimum sample size, the verification of the first type of error and the second type of error usually include statistical failure to run significance and additional test failure. For example, in linear regression, in addition to different linear regression algorithms, it involves goodness of fit tests, F tests, t tests, and confidence interval construction, all of which can be reversed by specifying the error rate and a priori sample characteristics to estimate the minimum sample size.

[0098] Considering that each factor item in the large model training theoretically conforms to the underlying assumptions of statistical analysis, the embodiment of the present application can propose a priori sample characteristics from the original large model, and then use the algorithm of the large model to reversely estimate the minimum number of samples (i.e., perform prior validation power analysis).

[0099] Therefore, in the embodiments of the present application, regardless of the underlying fitting and verification method, in practical applications, the statistical analysis quantity can be determined according to the following steps:

[0100] Step 1: Using the large model as a calibration template, obtain the preset statistical hypothesis error rate for the original learning data set (reflected in the large model as the built-in significance of the corresponding algorithm relationship, the test pass standard, or the lowest fit of all similar algorithms, etc.) and the prior sample characteristics (the statistical characteristics of the factor in the large model);

[0101] Step 2: Based on the statistical hypothesis error rate and prior sample characteristics, an algorithm or test function can be used to reverse engineer the number of learning samples required within the current correlation threshold range (i.e., the amount of statistical analysis required). Alternatively, a table can be created in advance and the number of learning samples can be obtained through a lookup table.

[0102] In the embodiment of the present application, the process of determining the statistical analysis value based on the statistical verification sample value table lookup method and the algorithm verification method can be described based on the following two aspects:

[0103] First, we can use the large model to determine the factor characteristics of the factor relationship and use the test and fitting algorithms to back-test. For example, for a single factor M with a specified factor relationship (1 / 0), we can use the T-test of the single factor M relationship. Assume that after multiple cluster analyses of the large model, factor M has a natural distribution characteristic. Assume that the following T-test formula is used to test whether the analysis passes the fit: T = (s^2 / (d^2*n))^1 / 2)

[0104] Among them, s is the standard deviation of the large model factor M, d is the difference between the average values ​​of M1 and M2, T is the internal assumption of the large model or the lowest value of the same type of regression of the model, and n is the quantity.

[0105] The T test is passed if the value is less than a certain value (e.g., T is less than 0.2). We can substitute the characteristic s value (standard deviation) and d value (mean value of M1 - mean value of M2) of the original factor M of the large model into the T value to obtain the minimum n value, which is the minimum number of samples required.

[0106] Second aspect: The embodiment of the present application can also determine the minimum required number of samples under the corresponding distribution characteristics by looking up the table in the appendix of the textbook (n value). That is, the embodiment of the present application can establish a table lookup for the minimum required number of samples for various factor relationships at the terminal for the purpose of estimating the n value.

[0107] After determining the statistical analysis amount under each correlation threshold range, that is, determining the minimum amount of learning data that can be extracted for each correlation threshold range, the original learning data set can be stratified and then the original learning data set can be screened from each stratified learning data set. This can be achieved by following the steps below:

[0108] Step 1: The factor items that covary or are correlated with the target factor items in the factor relationship table are regarded as covariant factor items, and the other factor items are regarded as random factor items;

[0109] Step 2: stratify the original learning data set according to the random factor item to obtain the stratified learning data set;

[0110] Step 3: Select learning materials whose total number is consistent with the minimum number of learning materials from the stratified learning material set.

[0111] Here, the original learning data set can be stratified according to the random factor items, so that the screened learning data can better meet the needs of actual model training, and then the learning data that matches the corresponding statistical analysis quantity can be extracted in each layer.

[0112] Taking the high-speed mobile safe driving target application scenario as an example, the minimum amount of learning materials to be extracted can be determined according to the following estimated demand table:

[0113] In summary, it can be determined that for the target factor item "raining", at least 162+81+81=324 learning data sets are required to learn it in model training. By determining the minimum number of learning data, the original learning data set can be screened to obtain the screened learning data set.

[0114] It can be seen that this application incorporates a factor relationship table containing prior knowledge to screen samples that meet the screening criteria. It can reduce and eliminate the problem that no matter what the algorithm is in artificial intelligence learning technology, it is essentially based on the algorithm for causal inference, and when too many non-white samples are included, it is likely to lead to causal inference errors, thereby being able to serve as AI learning reinforcement material.

[0115] In addition, considering that the causal inference algorithm can estimate the minimum number of samples required to be included as a factor, after screening the minimum number of samples, it can be used as supplementary material for AI learning, providing a basis for the perspective of processing experts, and further provided for learning use by AI terminals with limited computing power (such as automobiles).

[0116] It should be noted that in the related art, obtaining learning materials is time-consuming, labor-intensive and costly, and a small amount of important learning materials are difficult to obtain, resulting in errors or errors in learning results. The embodiment of the present application can effectively evaluate missing data and conduct more targeted collection through the estimated minimum sample size.

[0117] With respect to the method for generating an AI learning data set provided in the embodiment of the present application, the learning data set can also be adjusted in combination with actual application requirements.

[0118] First, based on the focused viewpoint images actually collected in the scene for which the trained model is applicable, the filtered learning data set can be expanded or replaced to obtain an updated learning data set. In this way, the determined updated learning data set will be processed into learning data that conforms to a certain perspective, so that the trained model will be more in line with the needs of the actual scene.

[0119] Secondly, the learning rules to be added can be obtained. The learning rules at least include the degree of correlation between the factor items to be added and other factor items. Then, based on the learning rules, the filtered learning data set is updated to obtain the updated learning data set. That is, factor items can be added to the learning data set at any time to meet more practical needs.

[0120] Third, the physical parameters of the terminal device for which the trained model is applicable can be obtained. Based on these physical parameters, the filtered learning data set can be updated to obtain an updated learning data set adapted to the terminal device. This updated learning data set will be more adapted to the actual needs of the terminal device. The physical parameters can include intrinsic and extrinsic parameters of the camera.

[0121] In order to better illustrate the practical application of the method for generating an AI learning data set provided in the embodiment of the present application, the following can be described in detail in combination with some actual application requirements of the terminal.

[0122] First, the terminal device transmits a certain requirement (such as driving in a certain area on rainy days, or a robot's work task in a certain factory), the algorithm on the device, and the device model to the host.

[0123] After the host confirms the demand, the host (can be one or more) selects the minimum learning data set (or supplementary data set) according to the demand, and forms the minimum learning data set here to achieve the main technical purpose of the embodiment of this application.

[0124] Finally, the minimum set of learning data can be reprocessed and returned to the terminal for use.

[0125] Based on the method for generating an AI learning data set provided in the above embodiment, the embodiment of the present application further provides a method for model training. Referring to FIG2 , which shows a flow chart of the method for model training provided in the embodiment of the present application, the method includes the following steps S201 to S202:

[0126] S201: Obtaining a filtered learning data set generated by a method for generating an AI learning data set;

[0127] S202: Perform at least one round of model training based on the filtered learning data set to obtain a trained model.

[0128] Here, one or more rounds of model training can be performed on the obtained filtered learning data set until a trained model suitable for the target application scenario is obtained. The one or more rounds of model training here can be used to train the correspondence between the learning data and the labeled factor items.

[0129] The trained model obtained based on the filtered learning data set can better adapt to the needs of the terminal device.

[0130] In the description of this specification, the description with reference to the terms "some possible embodiments", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application, and the above terms do not necessarily represent the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0131] About the method flow chart of the present application embodiment, some operations are described as the different steps performed in a certain order.Such flow chart belongs to illustrative and non-restrictive.Some steps described in this article can be grouped together and performed in a single operation, or some steps can be divided into multiple sub-steps and can be performed in an order different from that shown in this article.The various steps shown in the flow chart can be realized in any way by any circuit structure and / or tangible mechanism (for example, by the software, hardware (for example, the logical function realized by processor or chip) etc. running on computer equipment and / or its any combination).

[0132] Those skilled in the art will understand that, in the method described in the above specific embodiments, the writing order of each step does not mean a strict execution order, and the specific execution order of each step should be determined by its function and possible internal logic.

[0133] Based on the same inventive concept, an apparatus corresponding to the method is also provided in the embodiment of the present application. Since the principle of solving the problem by the apparatus in the embodiment of the present application is similar to the above-mentioned method in the embodiment of the present application, the implementation of the apparatus can refer to the implementation of the method, and the repeated parts will not be repeated.

[0134] 3 , which is a schematic diagram of a device for generating an AI learning data set provided in an embodiment of the present application, includes: an acquisition module 301 , a determination module 302 , and a screening module 303 ; wherein,

[0135] An acquisition module 301 is used to acquire an original learning data set, an algorithm, and a factor relationship table for a target application scenario; the factor relationship table includes multiple factor items adapted to the target application scenario and the relationships between the multiple factor items;

[0136] Determination module 302 is used to label the factors of each learning material in the original learning material set based on the algorithm, and determine the degree of correlation between the target factor item and other factor items required for learning in the model training process applicable to the target application scenario based on the factor relationship table;

[0137] The screening module 303 is used to screen the original learning data set based on the determined degree of association to obtain a screened learning data set. The screened learning data set is used as input data for model training in the target application scenario to generate a trained model.

[0138] The device for generating the above-mentioned AI learning data set can, when obtaining the original learning data set, calculation table and factor relationship table under the target application scenario, label the factor items of each learning data in the original learning data set based on the algorithm, and determine the degree of correlation between the target factor item and other factor items required for learning in the model training process applicable to the target application scenario based on the factor relationship table, and then filter the original learning data set based on the degree of correlation to obtain the filtered learning data set. The filtered learning data set obtained here is based on the degree of correlation between the key factor items (such as the target factor items) and other factor items determined by the model training, and the data is filtered. Since this method is based on the factor relationship table and thus has an automated feature, it will obtain the smallest possible data set with higher accuracy, which makes the filtered learning data set more orderly and adaptable to the training requirements of the target application scenario (such as terminal training requirements), and has better practicality and versatility.

[0139] Optionally, the screening module 303 is specifically configured to screen the original learning material set based on the determined correlation degree according to the following steps:

[0140] Determine the minimum amount of learning data required to learn the target factor item during model training for the target application scenario based on the degree of association;

[0141] The original learning data set is filtered according to the minimum number of learning materials to obtain a filtered learning data set.

[0142] Optionally, the screening module 303 is specifically configured to determine the minimum amount of learning materials required for the learning objective factor item according to the following steps:

[0143] In response to the correlation degree falling within a target correlation threshold range in the preset multi-level correlation threshold range, determining a statistical analysis amount within the target correlation threshold range;

[0144] Determine the minimum amount of learning materials based on statistical analysis.

[0145] Optionally, the screening module 303 is specifically configured to screen the original learning material set in the following steps:

[0146] Based on the factor items marked on each learning material in the original learning material set, target learning materials marked with target factor items are selected from the original learning material set; the number of target learning materials is the minimum number of learning materials required to learn the target factor items;

[0147] The filtered learning material set is determined based on the selected target learning materials.

[0148] Optionally, the screening module 303 is specifically configured to determine the statistical analysis amount within the target association threshold range according to the following steps:

[0149] Obtain the test conditions required for preliminary test analysis;

[0150] Determine whether each learning material within the target association threshold has passed the verification conditions based on a statistical verification sample value lookup table method or an algorithm verification method;

[0151] When it is determined that the test conditions are met, the statistical analysis quantity is determined.

[0152] Optionally, the screening module 303 is specifically configured to determine the statistical analysis amount according to the following steps:

[0153] Obtain the statistical hypothesis error rate and prior sample characteristics preset for the original learning data set;

[0154] The amount of statistical analysis is determined based on the statistical hypothesis error rate and prior sample characteristics.

[0155] Optionally, the screening module 303 is specifically configured to screen the original learning material set according to the following steps:

[0156] The factor items that covary or are correlated with the target factor items in the factor relationship table are regarded as covariant factor items, and the other factor items are regarded as random factor items;

[0157] Stratify the original learning data set according to the random factor item to obtain the stratified learning data set;

[0158] Select learning materials from the stratified learning material set whose total number is consistent with the minimum number of learning materials.

[0159] Optionally, the device further comprises:

[0160] The first updating module 304 is used to obtain the focused viewpoint image collected in the scene applicable to the trained model; expand or replace the filtered learning data set based on the focused viewpoint image to obtain an updated learning data set.

[0161] Optionally, the device further comprises:

[0162] The second updating module 305 is used to obtain the learning rules to be added; the learning rules at least include the correlation between the characteristics of the factor item to be added and other factor items; based on the learning rules, the filtered learning data set is updated to obtain an updated learning data set.

[0163] Optionally, the device further comprises:

[0164] The third updating module 306 is used to obtain the physical parameters of the terminal device to which the trained model is applicable; and update the screened learning data set based on the physical parameters to obtain an updated learning data set adapted to the terminal device.

[0165] 4 , which is a schematic diagram of a model training device provided in an embodiment of the present application, includes: an acquisition module 401 and a training module 401; wherein,

[0166] An acquisition module 401 is used to acquire a filtered learning data set generated based on a method for generating an AI learning data set;

[0167] The training module 402 is used to perform at least one round of model training based on the filtered learning data set to obtain a trained model.

[0168] It should be noted that the device in the embodiment of the present application can implement each process of the embodiment of the aforementioned method and achieve the same effects and functions, which will not be repeated here.

[0169] Corresponding to the method for generating an AI learning data set in FIG1 or the method for model training in FIG2 , an embodiment of the present application further provides an electronic device, as shown in FIG5 , which is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, including:

[0170] Processor 51, memory 52, and bus 53; memory 52 is used to store execution instructions, including internal memory 521 and external memory 522; the internal memory 521 is also called internal memory, which is used to temporarily store the operation data of the processor 51 and the data exchanged with the external memory 522 such as a hard disk. The processor 51 exchanges data with the external memory 522 through the internal memory 521. When the electronic device 500 is running, the processor 51 and the memory 52 communicate through the bus 53, so that the processor 51 executes the following instructions:

[0171] Obtaining the original learning data set, algorithm, and factor relationship table for the target application scenario; the factor relationship table includes multiple factor items adapted to the target application scenario and the correlation relationship between the multiple factor items;

[0172] Based on the algorithm, each learning material in the original learning data set is labeled with factors, and the degree of correlation between the target factor item and other factor items required for learning in the model training process suitable for the target application scenario is determined based on the factor relationship table;

[0173] The original learning data set is screened based on the determined degree of association to obtain a screened learning data set, which is used as input data for model training in the target application scenario to generate a trained model;

[0174] Or the processor 51 executes the following instructions:

[0175] Obtaining a filtered learning data set generated by a method for generating an AI learning data set;

[0176] Perform at least one round of model training based on the filtered learning data set to obtain a trained model.

[0177] The present application also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the steps of the method described in the above method embodiment. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0178] An embodiment of the present application also provides a computer program product, which carries program code. The instructions included in the program code can be used to execute the steps of the method described in the above method embodiment. For details, please refer to the above method embodiment and will not be repeated here.

[0179] The computer program product may be implemented in hardware, software, or a combination thereof. In one embodiment, the computer program product is implemented as a computer storage medium. In another embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0180] Each embodiment of this application is described in a progressive manner. The same or similar parts between the various embodiments can be referenced in each other, and each embodiment focuses on the differences from other embodiments. In particular, the descriptions of the device, equipment, and computer-readable storage medium embodiments are simplified because they are basically similar to the method embodiments. For relevant parts, please refer to the partial description of the method embodiments.

[0181] The apparatus, equipment and computer-readable storage medium provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the apparatus, equipment and computer-readable storage medium also have similar beneficial technical effects to their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the apparatus, equipment and computer-readable storage medium will not be repeated here.

[0182] It should be understood by those skilled in the art that the embodiments of the present application can be implemented as methods and apparatus (devices or systems), or computer-readable storage media. Therefore, the present application can adopt a complete hardware implementation, a complete software implementation, or an implementation combining software and hardware. Moreover, the present application can take the form of a computer-readable storage medium implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, read-only compact disk storage (CD-ROM), optical storage, etc.) containing computer-usable program code.

[0183] The present application is described with reference to the flowchart and / or block diagram of the method, device (equipment or system) and computer-readable storage medium according to the embodiment of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram and the combination of the process and / or box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processing machine or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the function specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0184] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, wherein the instruction device implements the functions specified in one or more processes in the flowchart and / or one or more blocks in the block diagram.

[0185] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0186] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0187] Memory can include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash memory (Flash RAM). Memory is an example of a computer-readable medium.

[0188] Computer-readable media include permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory, read-only memory, electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. In addition, although the operations of the method of the present application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in this specific order, or that all shown operations must be performed to achieve the desired result. In addition, certain steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple substeps.

[0189] Although the spirit and principles of the present application have been described above with reference to several specific embodiments, it should be understood that the present application is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features of these aspects cannot be combined. The present application is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for generating an AI learning data set, comprising: Obtain the original learning data set, algorithm, and factor relationship table for the target application scenario; The factor relationship table includes multiple factor items adapted to the target application scenario and the association relationships between the multiple factor items; Labeling each learning material in the original learning material set with factor items based on the algorithm, and determining, based on the factor relationship table, the degree of correlation between the target factor item to be learned and other factor items in the model training process applicable to the target application scenario; The original learning data set is screened based on the determined degree of association to obtain a screened learning data set, and the screened learning data set is used as input data for model training in the target application scenario to generate a trained model.

2. The method according to claim 1, wherein The screening of the original learning data set based on the determined association degree includes: Determining, based on the degree of association, a minimum amount of learning materials required to learn the target factor item during the model training process of the target application scenario; The original learning material set is screened according to the minimum number of learning materials to obtain a screened learning material set.

3. The method according to claim 2, wherein: Follow these steps to determine the minimum amount of learning material required to learn the target factor item: In response to the correlation degree belonging to a target correlation threshold range in a preset multi-level correlation threshold range, determining a statistical analysis amount within the target correlation threshold range; The minimum amount of learning materials is determined based on the statistical analysis amount.

4. The method according to claim 3, wherein: The screening of the original learning material set according to the minimum number of learning materials includes: Based on the factor items marked on each learning material in the original learning material set, target learning materials marked with target factor items are selected from the original learning material set; the number of target learning materials is the minimum number of learning materials required to learn the target factor items; The filtered learning material set is determined based on the selected target learning material.

5. The method according to claim 3 or 4, wherein: The determining of the statistical analysis amount within the target association threshold range includes: Obtain the test conditions required for preliminary test analysis; Determining whether each learning material within the target association threshold range passes the verification condition based on a statistical verification sample value table lookup method or an algorithm verification method; When it is determined that the test condition is met, the statistical analysis amount is determined.

6. The method according to claim 5, wherein: Determining the statistical analysis amount includes: Obtaining a statistical hypothesis error rate and a priori sample characteristics preset for the original learning data set; The statistical analysis quantity is determined based on the statistical hypothesis error rate and the prior sample characteristics.

7. The method according to any one of claims 2 to 6, wherein The screening of the original learning material set according to the minimum number of learning materials includes: The factor items in the factor relationship table that are covariant or correlated with the target factor item are used as covariant factor items, and the other factor items are used as random factor items; stratifying the original learning data set according to the random factor item to obtain a stratified learning data set; Select learning materials whose total amount is consistent with the minimum amount of learning materials from the stratified learning material set.

8. The method according to any one of claims 1 to 7, wherein The method further comprises: Obtaining a focused viewpoint image captured in a scenario applicable to the trained model; The filtered learning data set is expanded or replaced based on the focused viewpoint image to obtain an updated learning data set.

9. The method according to any one of claims 1 to 8, wherein The method further comprises: Acquire a learning rule to be added; the learning rule at least includes the correlation degree between the factor item to be added and other factor items; The filtered learning data set is updated based on the learning rule to obtain an updated learning data set.

10. The method according to any one of claims 1 to 9, wherein The method further comprises: Obtaining physical parameters of the terminal device to which the trained model is applicable; The filtered learning data set is updated based on the physical parameters to obtain an updated learning data set adapted to the terminal device.

Images