CN114419420A

CN114419420A - Model detection method, device, equipment and storage medium

Info

Publication number: CN114419420A
Application number: CN202111663170.4A
Authority: CN
Inventors: 唐晓璇; 吕广奕; 章学敏
Original assignee: Lenovo Beijing Ltd
Current assignee: Lenovo Beijing Ltd
Priority date: 2021-12-31
Filing date: 2021-12-31
Publication date: 2022-04-29

Abstract

The embodiment of the application discloses a model detection method, a device, equipment and a storage medium, wherein the method comprises the following steps: inputting data to be detected into a target prediction model to obtain a prediction result; extracting features of data to be detected, and determining a plurality of features corresponding to the data to be detected; performing multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected; determining training sample distribution information corresponding to a target prediction model through a measurement model, and performing measurement calculation on the training sample distribution information and target characteristics to obtain a measurement result; and determining whether the prediction result is correct or not according to the measurement result. Therefore, whether the prediction result of the target prediction model is correct or not can be judged by using the measurement result between the target characteristic of the data to be detected and the distribution information of the training sample, and the target prediction model in the production environment is supervised.

Description

Model detection method, device, equipment and storage medium

Technical Field

The present application relates to the field of computer software technologies, and in particular, to a model detection method, apparatus, device, and storage medium.

Background

With the continuous development of artificial intelligence technology, various prediction models are applied to industrial application scenarios. For the prediction model, when a sample to be detected similar to the training sample is encountered, the accuracy rate is higher; however, when an unknown sample different from the training sample is encountered, the accuracy of the prediction model is sharply reduced, so that the application requirement of a real scene cannot be met.

Disclosure of Invention

The application provides a model detection method, a model detection device and a storage medium.

The technical scheme of the application is realized as follows:

in a first aspect, an embodiment of the present application provides a model detection method, where the method includes:

inputting data to be detected into a target prediction model to obtain a prediction result;

extracting features of data to be detected, and determining a plurality of features corresponding to the data to be detected;

performing multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected;

determining training sample distribution information corresponding to a target prediction model through a measurement model, and performing measurement calculation on the training sample distribution information and target characteristics to obtain a measurement result;

and determining whether the prediction result is correct or not according to the measurement result.

In a second aspect, an embodiment of the present application provides a model detection apparatus, including:

the prediction unit is configured to input the data to be detected into the target prediction model to obtain a prediction result;

the extraction unit is configured to extract features of the data to be detected and determine a plurality of features corresponding to the data to be detected;

the fusion unit is configured to perform multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected;

the calculation unit is configured to determine training sample distribution information corresponding to the target prediction model through the measurement model, and perform measurement calculation on the training sample distribution information and the target characteristics to obtain a measurement result;

and the processing unit is configured to determine whether the prediction result is correct according to the measurement result.

In a third aspect, an embodiment of the present application provides a model detection apparatus, which includes a memory and a processor; wherein the content of the first and second substances,

a memory for storing a computer program capable of running on the processor;

a processor for performing the steps of the method according to the first aspect when running the computer program.

In a fourth aspect, the present application provides a computer storage medium storing a computer program which, when executed, implements the steps of the method according to the first aspect.

The embodiment of the application provides a model detection method, a device, equipment and a storage medium, wherein the method comprises the following steps: inputting data to be detected into a target prediction model to obtain a prediction result; extracting features of the data to be detected, and determining a plurality of features corresponding to the data to be detected; performing multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected; determining training sample distribution information corresponding to the target prediction model through a measurement model, and performing measurement calculation on the training sample distribution information and the target characteristics to obtain a measurement result; and determining whether the prediction result is correct or not according to the measurement result. Therefore, by carrying out feature extraction and multi-mode fusion processing on the data to be detected, the obtained target features can more comprehensively reflect the relevant information of the data to be detected, and the accuracy of subsequent judgment is improved; in addition, whether the data to be detected is abnormal is determined by using a measurement result between the target characteristic of the data to be detected and the distribution information of the training sample, so that whether the prediction result of the target prediction model is correct is judged, the target prediction model in a production environment can be supervised, and the wrong prediction result is reduced.

Drawings

Fig. 1 is a schematic flowchart of a model detection method according to an embodiment of the present disclosure;

fig. 2 is a schematic flowchart of another model detection method provided in an embodiment of the present application;

fig. 3 is a schematic structural diagram of a model detection apparatus according to an embodiment of the present disclosure;

FIG. 4 is a schematic diagram illustrating an operation process of a model detection apparatus provided in an embodiment of the present application;

FIG. 5 is a schematic structural diagram of another model detection apparatus provided in an embodiment of the present application;

fig. 6 is a schematic hardware structure diagram of an electronic device according to an embodiment of the present disclosure;

fig. 7 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant application and are not limiting of the application. It should be noted that, for the convenience of description, only the parts related to the related applications are shown in the drawings.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to be limiting of the application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.

It should be noted that the terms "first \ second \ third" are used merely to distinguish similar objects and do not represent a specific ordering for the objects, and it should be understood that "first \ second \ third" may be interchanged under certain ordering or sequence if so permitted so that the embodiments of the present application described herein can be implemented in other orders than that shown or described herein.

It should be appreciated that as artificial intelligence evolves rapidly, various predictive models (e.g., edge inference models) continue to fall on the floor of actual production. For the prediction model, a plurality of model parameters need to be determined through a large number of training samples, and the model parameters can be used in the actual production process. However, in the actual production process, some unknown samples having a large difference from the training samples may be generated due to the change of the production environment, and at this time, the prediction model cannot accurately judge the unknown samples, so that the accuracy of the prediction model is sharply reduced, and even the accuracy of the prediction model cannot reach the industrial application standard. That is, the unknown samples can seriously affect the prediction accuracy when the model is applied, and cause problems to the actual production process.

Based on this, the embodiment of the present application provides a model detection method, and the basic idea is: inputting data to be detected into a target prediction model to obtain a prediction result; extracting features of the data to be detected, and determining a plurality of features corresponding to the data to be detected; performing multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected; determining training sample distribution information corresponding to the target prediction model through a measurement model, and performing measurement calculation on the training sample distribution information and the target characteristics to obtain a measurement result; and determining whether the prediction result is correct or not according to the measurement result. Therefore, by carrying out feature extraction and multi-mode fusion processing on the data to be detected, the obtained target features can more comprehensively reflect the relevant information of the data to be detected, and the accuracy of subsequent judgment is improved; in addition, whether the data to be detected is abnormal is determined by using a measurement result between the target characteristic of the data to be detected and the distribution information of the training sample, so that whether the prediction result of the target prediction model is correct is judged, the target prediction model in a production environment can be supervised, and the loss caused by the wrong prediction result is reduced.

Embodiments of the present application will be described in detail below with reference to the accompanying drawings.

In an embodiment of the present application, referring to fig. 1, a flowchart of a model detection method provided in an embodiment of the present application is shown. As shown in fig. 1, the method may include:

s101: and inputting the data to be detected into the target prediction model to obtain a prediction result.

It should be noted that the model detection method provided in the embodiment of the present application may be applied to various electronic devices having a computing function.

In the embodiment of the present application, the target prediction model may be an inference model in various fields, such as an image processing field, a speech recognition field, a natural language recognition field, and the like. Specifically, the target prediction model is used for performing specific detection on data to be detected so as to output a prediction result.

S102: and performing feature extraction on the data to be detected, and determining a plurality of features corresponding to the data to be detected.

It should be noted that the feature extraction is performed through multiple dimensions, so as to obtain multiple features corresponding to the data to be detected. Therefore, the information of the data to be detected under different dimensionalities can be reflected through a plurality of characteristics, and the accuracy of subsequent judgment is improved.

S103: and performing multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected.

It should be noted that, through the multi-modal fusion process, features in multiple dimensions are fused as target features, so as to facilitate subsequent processing. Here, the multimodal fusion process may employ various principles, such as front-end fusion, middle fusion, and back-end fusion.

S104: and determining training sample distribution information corresponding to the target prediction model through the measurement model, and performing measurement calculation on the training sample distribution information and the target characteristics to obtain a measurement result.

The training sample distribution information is a feature of sample data (hereinafter, simply referred to as an original training sample) used for training the target prediction model.

In a specific embodiment, the training sample distribution information is obtained by clustering the respective target features of the original training samples. At this time, the target features of the original training samples form a plurality of cluster centers after clustering, and then the measurement result can be obtained through the distance between the target features and the cluster centers.

It should be understood that the target prediction model is obtained after training with a large number of original training samples. If the data to be detected is similar to the original training sample, the prediction result is reliable; if the deviation between the data to be detected and the original training sample is large, the prediction result of this time is also unreliable. Based on the method, the measurement model obtains a measurement result by detecting the distance between the data to be detected and the original training sample. Therefore, whether the prediction result of the target prediction model is correct or not can be further determined according to the measurement result.

Further, in some embodiments, the metric model may include an anti-forgetting module and at least one metric calculation submodel, where the determining, by the metric model, training sample distribution information corresponding to the target prediction model, and performing metric calculation on the training sample distribution information and the target feature to obtain a metric result includes:

acquiring training sample distribution information from an anti-forgetting module;

calculating target characteristics and training sample distribution information based on at least one measurement calculation sub-model to obtain at least one measurement score information;

at least one metric score information is determined as a metric result.

It should be noted that the metric model includes an anti-forgetting module for storing the distribution information of the training samples and a metric calculation submodel for performing metric calculation.

The measurement calculation sub-model can perform measurement calculation and scoring processing on the target characteristics and the original training sample distribution characteristics so as to obtain measurement score information. Thus, by measuring the score information, the similarity between the data to be detected and the original training sample can be obtained.

The number of the measurement calculation submodels is at least one, so that the similarity between the data to be detected and the original training sample can be judged from different dimensions, and the comprehensiveness and the accuracy of the measurement result are improved.

S104: and determining whether the prediction result is correct or not according to the measurement result.

It should be noted that, if the measurement result indicates that the similarity between the data to be detected and the original training sample is higher, the prediction result is determined to be correct; if the metric indicates that the similarity between the data to be detected and the original training sample is low, it is determined that the prediction is erroneous.

Therefore, by the model detection method provided by the embodiment of the application, whether the prediction result of the target prediction model is correct or not can be judged according to the difference between the sample to be detected and the original training sample, and the automatic detection of the prediction result is realized.

Further, in some embodiments, the method may further comprise:

under the condition that the prediction result is correct, outputting the prediction result;

determining the data to be detected as abnormal data under the condition that the prediction result is wrong; storing the abnormal data in a target area; and when the number of the abnormal data in the target area reaches a target threshold value, updating the target prediction model and the measurement model by using all the abnormal data.

Therefore, when the prediction result is judged to be correct, the prediction result is output to the user so that the prediction result can be applied in the next step, when the prediction result is judged to be wrong, the data to be detected is stored as abnormal data, and the target prediction model is updated by using the abnormal data subsequently. Therefore, on one hand, the credibility of the target prediction model can be improved, and negative influence on the production process caused by wrong prediction results is avoided; on the other hand, the updating direction of the target prediction model is clarified, and the model updating efficiency is improved. In another aspect, with the model detection method provided by the embodiment of the present application, if the abnormal data is accumulated to a certain amount (or reaches other model update conditions), the target prediction model can be automatically updated by using the abnormal data in the target region, and interaction with a user is not required, so that an imperceptible update is realized.

After the target prediction model is updated with the abnormal data, the abnormal data actually becomes a part of the "original training samples" of the target prediction model, and therefore, it is necessary to update the metrology model with the abnormal data to update the training sample distribution information in the metrology model.

Further, the anomaly data is divided into three anomaly categories so as to facilitate the subsequent selection of different strategies to update the target prediction model, and therefore, in some implementations, the method may further include:

performing interpolation calculation on at least one measurement score information to obtain an abnormal value of abnormal data;

and comparing the abnormal value of the abnormal data with the abnormal threshold value to determine the abnormal type of the abnormal data.

It should be noted that, an abnormal value is calculated by using at least one metric score information, and the higher the abnormal value is, the greater the deviation of the abnormal data from the original training sample is. Here, the interpolation calculation is not essential, and the abnormal value may be determined by a method such as a mean calculation.

It should be noted that, in the embodiment of the present application, the abnormal data is classified into three categories according to the deviation degree between the abnormal data and the original training sample: new type samples, difficult to learn samples, or drift samples. The sample difficult to learn means that the abnormal data belongs to the coverage range of the original training sample, but the prediction is difficult to perform; the drift sample means that the abnormal data has a deviation from the original training sample, but the deviation is not large; the new type sample means that the abnormal data is of a different type from the original training sample.

For example, assuming that the target prediction model is used to identify cats in the data to be detected, and the original training samples are related pictures of a large number of citrus cats, the learning samples, the drift samples, and the new type samples can be understood as follows: the sample difficult to learn is still a mandarin cat picture, and is different from the original training sample; the drift sample may be a black cat picture; the new category sample may be a dog picture.

In a specific embodiment, the outliers of the difficult-to-learn sample, the drift sample, and the new type of sample increase in sequence. Correspondingly, the abnormity threshold comprises a first abnormity threshold and a second abnormity threshold, and the first abnormity threshold is smaller than the second abnormity threshold;

determining abnormal data as a difficult-to-learn sample under the condition that the abnormal value of the abnormal data is less than or equal to a first abnormal threshold;

determining the abnormal data as a drift sample under the condition that the abnormal value of the abnormal data is larger than the first abnormal data and the abnormal value of the abnormal data is smaller than or equal to a second abnormal threshold;

and determining the abnormal data as a new type sample under the condition that the abnormal value of the abnormal data is larger than the abnormal data of the first person.

Therefore, through the steps, which type of abnormal category the abnormal data belongs to can be determined, so as to guide the user to label in the following.

In some embodiments, after determining the anomaly category of the anomalous data, the method may further include determining annotation suggestion information for the anomalous data according to the anomaly category of the anomalous data;

sending abnormal data to a user according to the labeling suggestion information of the abnormal data;

and receiving the real label of the abnormal data sent by the user.

Note that the annotation suggestion information is information for guiding the user to perform annotation. For example, in order to facilitate the user to mark the abnormal data quickly, some preset options may be provided for the user to select. The annotation suggestion information is what preset options are provided.

For example, for an intractable sample and a drift sample, the user needs to be asked whether the prediction result is wrong; for new types of samples, the user needs to be asked what the true category of the anomalous data is.

It should be noted that, since the samples that are difficult to learn, the drift samples, and the new type samples are automatically classified according to the abnormal threshold, a classification error may occur. Thus, the anomaly threshold can be designed to be an adaptive value. In other words, after the user marks the real label of the abnormal data, the abnormal threshold is also adaptively adjusted according to the real label of the abnormal data.

In some embodiments, as shown in fig. 2, the updating the target prediction model with all the abnormal data may include:

s201: and counting the abnormal data belonging to the new type of samples, the abnormal data belonging to the samples difficult to learn and the abnormal data belonging to the drift samples to obtain a statistical result.

It should be noted that, when the number of abnormal data in the target area reaches the target threshold, it indicates that the reliability of the target prediction model cannot meet the requirements of the production practice, and therefore needs to be updated.

After the updating is started, counting the abnormal type of the abnormal data so as to determine the updating direction of the target prediction model.

S202: and determining the data deviation value according to the statistical result.

It should be noted that, here, the data offset value is used to indicate the degree of deviation between the data to be detected and the original training sample. In other words, the deviation degree between the data to be detected and the original training sample can be quantitatively detected through the data deviation value.

According to the foregoing, the degree of offset of each of the difficult-to-learn sample, the drift sample, and the new-type sample from the original training sample is different. Illustratively, if the abnormal data is mainly the hard-to-learn sample and/or the drift sample, the data deviation value is small, which indicates that the deviation between the data to be detected and the original training sample in the production practice is small, and if the abnormal data is mainly the new type sample, the data deviation value is large, which indicates that the deviation between the data to be detected and the original training sample in the production practice is large.

Here, a specific calculation method of the data offset value may be determined according to an actual application scenario. For example, the proportion of the new type sample to all the abnormal data can be used as the data offset value; or, determining a certain weight value for each abnormal data, and then performing weighted calculation.

S203: and under the condition that the data deviation value is larger than the deviation threshold value, updating the target prediction model by using all abnormal data and all real labels of the abnormal data based on the first type of updating strategy.

S204: and under the condition that the data deviation value is smaller than or equal to the deviation threshold value, updating the target prediction model by using all abnormal data and all real labels of the abnormal data based on the second type of updating strategy.

Here, the retention ratio of the original model features of the target prediction model in the first class of update strategies is smaller than the retention ratio of the original model features of the target prediction model in the second class of update strategies.

It should be noted that, if the data deviation value is larger, it indicates that the deviation between the data to be detected and the original training sample in the production practice is larger, the importance of the original model features is smaller, and at this time, fewer original model features are reserved in the updating process; on the contrary, if the data deviation value is smaller, it indicates that the deviation between the data to be detected and the original training sample in the production practice is smaller, the importance of the original model features is larger, and more original model features are retained in the updating process.

Illustratively, the first type of update policy may be Few-Shot Learning, and the second type of update policy may be Transfer Learning.

That is to say, the target prediction model is obtained by training the neural network model by using the original training samples, the target prediction model mostly depends on the original training samples selected in the initial model training process, and the target prediction model can only be applied to the data to be detected, which are similar to the original training samples. However, in an actual production environment, data to be detected is constantly changed, and it cannot be guaranteed that an original training sample can cover various types of data to be detected, so that the target prediction model needs to be constantly updated to adapt to the constantly changing data.

In particular, the purpose of model update is to adapt to the data to be detected (i.e. abnormal data) with large deviation from the original training sample, so the specific strategy of model update also depends on the type of abnormal data. In the embodiment of the present application, there are at least the following three scenarios:

scene one: the data to be detected in the production environment basically has no change, and only some samples which are difficult to learn are added, for example, abnormal data still belong to the coverage range of the original training sample, but the result of the target prediction model is still wrong.

Scene two: compared with the original training sample, the data to be detected in the production environment has some deviation, but the overall deviation is not large, namely the sample drifts, for example, the original training data is orange cat, and the current training data is black cat.

Scene three: compared with the original training sample, the data to be detected in the production environment has larger deviation, and a new type of sample already appears, for example, the original training data is a cat, and a dog exists in the sample.

For the above situations, different update strategies need to be adopted during model update: (1) for the first scene and the second scene, a Transfer Learning mode is adopted, namely a second type of updating strategy; (2) for scene three, Few-Shot Learning is used. The general model updating method is to deduce the characteristics of the original data first and update the characteristics during updating.

The difference between the two updating modes of Transfer Learning and Few-Shot Learning lies in the different ratio of the pre-stored features. That is to say, the update strategy completely depends on the offset degree of the input data, but the model detection method provided by the embodiment of the present application can quantitatively detect the offset degree of the input data, and the two different update strategies are selected according to the detected different offset degrees when the model is updated.

In brief, the updating strategy completely depends on the offset degree of the input data, the method can quantitatively detect the offset degree of the input data, and the two different updating strategies are selected according to different detected offset degrees during model updating.

In this way, the updating strategy of the target prediction model can be clarified according to the abnormal type of the abnormal data, so that the model updating efficiency is improved.

Further, in some embodiments, the method may further comprise:

training an original training sample, and establishing a target prediction model; and the number of the first and second groups,

acquiring target characteristics of an original training sample, and determining training sample distribution information of a target prediction model according to the target characteristics of the original training sample;

storing the distribution information of the training samples to an anti-forgetting module;

under the condition that the prediction result is correct, updating the distribution information of the training samples in the anti-forgetting module by using the target characteristics of the data to be detected;

and under the condition that the prediction result is wrong, determining the target characteristics of the abnormal data to update the distribution information of the training samples in the anti-forgetting module in the process of training the target prediction model by using the abnormal data and the real labels of the abnormal data.

It should be noted that the target features of the original training samples and the target features of the abnormal data may be calculated by referring to the calculation method of the target features of the data to be detected.

It should be noted that the anti-forgetting module can be understood as a database storing group feature vectors capable of expressing all input samples, wherein each specific feature vector can represent a part of types of data samples. The anti-forgetting module is used for preventing catastrophic forgetting in the model updating process. Catastrophic forgetting means that the target prediction model forgets the original data features while learning new features, resulting in poor performance on the initial original task. When the model is updated, it is easier to train the model with new data (i.e. abnormal data in the embodiment of the present application) and let the model learn the characteristics of the new data, so that the model can be well represented on the new data with higher accuracy, and it is more difficult to learn the new characteristics while not forgetting the characteristics of the historical data, and the model can be highly accurate on both new tasks and old tasks. The anti-forgetting module prevents catastrophic forgetting by updating only a small portion of the population feature vectors most relevant to new data and controlling the degree of updating of these vectors.

In addition, the generation of the anti-forgetting module can be divided into the following steps: (1) in the initialization stage, namely when the target prediction model is deployed on the equipment, the anti-forgetting module is calculated according to all original training samples in the training set. It should be appreciated that the original training samples are, for the most part, correctly predicted by the target prediction model. (2) in the model using stage, updating the target characteristics of the sample to be detected with correct prediction results into an anti-forgetting module; (2) and in the model updating stage, the target characteristics of the abnormal samples are updated to the anti-forgetting module.

In addition, the essence of the measurement model is a nonparametric model, and the measurement model can be updated by updating the training data distribution information in the measurement model, so that the calculation speed is high, and the calculation pressure is low.

In summary, the model detection method utilizes the multi-modal fusion features and various learnable metric models to detect data to be detected from different dimensions, can determine whether a sample to be detected is abnormal data, and can also give classification suggestions of the abnormal data; in addition, the measurement model is a nonparametric model, the output of the detected target prediction model can be corrected, the calculation amount is small, and extra pressure cannot be brought to equipment. Therefore, by the model detection method provided by the embodiment of the application, the prediction result of the target detection model can be judged secondarily, the prediction accuracy of the target detection model is improved, the updating direction of the target detection model can be determined, and the updated target detection model is suitable for the actual production environment.

The embodiment of the application provides a model detection method, which comprises the following steps: inputting data to be detected into a target prediction model to obtain a prediction result; extracting features of data to be detected, and determining a plurality of features corresponding to the data to be detected; performing multi-mode fusion processing on the plurality of characteristics to obtain target characteristics corresponding to the data to be detected; determining training sample distribution information corresponding to a target prediction model through a measurement model, and performing measurement calculation on the training sample distribution information and target characteristics to obtain a measurement result; and determining whether the prediction result is correct or not according to the measurement result. Therefore, by carrying out feature extraction and multi-mode fusion processing on the data to be detected, the obtained target features can more comprehensively reflect the relevant information of the data to be detected, and the accuracy of subsequent judgment is improved; in addition, whether the data to be detected is abnormal is determined by using a measurement result between the target characteristic of the data to be detected and the distribution information of the training sample, so that whether the prediction result of the target prediction model is correct is judged, the target prediction model in a production environment can be supervised, and the loss caused by the wrong prediction result is reduced.

In another embodiment of the present application, refer to fig. 3, which shows a schematic structural diagram of a model detection apparatus 30 provided in an embodiment of the present application. As shown in fig. 3, the model detection apparatus may include an abnormal value detection module 301, a user annotation suggestion module 302, and an abnormal sample management module 303.

Referring to fig. 4, a schematic diagram of an operation process of a model detection apparatus 30 according to an embodiment of the present application is shown. As shown in fig. 4, after the target prediction model predicts the data to be detected and obtains the model prediction result, the model detection apparatus 30 further determines whether the model prediction result is correct through the abnormal value detection module 301, the user labeling suggestion module 302 and the abnormal sample management module 303. The concrete description is as follows:

the working process of the abnormal value detection module 301 includes: (1) the pre-training network extracts features. And after receiving the data to be detected, extracting the characteristics of the data to be detected by using the pre-trained network. (2) And (5) multi-modal fusion treatment. Due to the diversity of samples in an actual scene, the embodiment of the application performs multi-mode fusion processing on various characteristics of the data to be detected, so as to obtain the target characteristics of the data to be detected. (3) And carrying out metric calculation by utilizing the training data distribution information. And acquiring training data distribution information from the anti-forgetting module, wherein the training data distribution information is used for indicating the characteristics of a training sample of the target prediction model, and performing measurement calculation on the target characteristics of the data to be detected and the training data distribution information under different dimensions to obtain various measurement results. (4) And comprehensively scoring based on the various measurement results. And (4) integrating the scoring conditions of the various measurement results, and calculating whether the prediction result of the target prediction model is correct or not.

If the prediction result is correct, the prediction result is output, and the training sample distribution information in the anti-forgetting module is updated by using the target characteristics of the data to be detected, so that the accuracy of judging subsequent samples is improved, and 'catastrophic forgetting' is prevented. And if the prediction result is wrong, comprehensively calculating the abnormal value of the data to be detected by using an interpolation algorithm according to the scoring information of the various measurement results. Subsequently, the user interaction labeling module 302 may also give an abnormal category suggestion according to the task scenario and the abnormal value of the data to be detected, and push the data to be detected to the user interaction interface to guide the user to label the data to be detected, if the data to be detected is a new category sample.

Specifically, the working process of the user interaction annotation module 302 includes: according to the comparison result between the abnormal value of the data to be detected calculated by the abnormal value detection module 301 and the adaptive abnormal threshold, the abnormal data belonging to one of the three abnormal samples is distinguished, and then manual class marking and verification are guided. Here, the samples classified as erroneous can be classified into three types: the first is a hard-to-learn sample, the second is a concept drift sample, and the third is a sample distributed far away from the training data set.

In this way, after the user interaction labeling module 302 determines the abnormal category of the abnormal data, a labeling suggestion is given according to the abnormal category of the abnormal data, so as to guide the user to label the abnormal data, that is, "user interaction labeling abnormal data" in fig. 4.

The abnormal sample data management module 303 may collect, manage, and store the abnormal data labeled by the user, which is equivalent to the target area. Then, a dynamic update strategy is formulated according to the type and the quantity of the abnormal data accumulated in the target area, the detection type of the target prediction model, the specific production scene and the like, the model update process of the target prediction model is automatically triggered subsequently, the detection capability of the target prediction model in the current production environment is improved, and the model accuracy is improved, namely, the model is updated according to data information in fig. 4.

In addition, in the process of updating the model, not only the target prediction model needs to be updated by using the data information in the abnormal sample data management and storage module 303, but also the anti-forgetting module needs to be updated.

In summary, in an actual industrial application scenario, due to a change in a production environment and the like, the target prediction model encounters an unknown sample (equivalent to the above abnormal data) which cannot be determined, so that the model accuracy is rapidly reduced, and even the model accuracy cannot meet an industrial application standard. The unknown samples can seriously affect the prediction accuracy when the model is applied, so that the detection of abnormal data affecting the model accuracy in an actual scene is a core problem, and the step is important for subsequent model updating and lifelong learning.

Based on the problem, the embodiment of the application provides a model detection device based on metric calculation, which has the following advantages: on one hand, the embodiment of the application uses the learnable measurement model (equivalent to a measurement learning submodel) with various dimensions to comprehensively score various measurement results, so that the model detection device can detect various types of abnormal data, is more suitable for a real production environment, and improves the accuracy of the model. On the other hand, the model detection device detects the abnormal value of the abnormal data and compares the abnormal value of the abnormal data with the adaptive abnormal threshold value to give the abnormal category (including an intractable sample, a drift sample or a new type sample) of the abnormal data and guide the user to label the abnormal data, which is important for selecting a model updating strategy subsequently. In yet another aspect, the model detection apparatus is a non-parametric model that does not require the consumption of significant computational resources to train the model, which also allows the apparatus to be updated on the fly.

The model detection device provided by the application can at least partially improve the following problems: detecting data to be detected of the target prediction model in an actual production environment, thereby detecting abnormal data which are possibly predicted wrongly by the target prediction model, and simultaneously determining a real label for the detected abnormal data; abnormal data are collected and stored, updating of the target prediction model is automatically triggered, and the prediction precision of the target prediction model on the abnormal data is improved; meanwhile, the model detection device can also be adapted to a self-learning dynamic updating system.

The embodiment of the application provides a model detection method, and the specific implementation method of the embodiment is elaborated in detail through the embodiment, so that the obtained target characteristics can more comprehensively reflect the relevant information of the data to be detected by performing characteristic extraction and multi-mode fusion processing on the data to be detected, and the accuracy of subsequent judgment is improved; in addition, whether the data to be detected is abnormal is determined by using a measurement result between the target characteristic of the data to be detected and the distribution information of the training sample, so that whether the prediction result of the target prediction model is correct is judged, the target prediction model in a production environment can be supervised, and the loss caused by the wrong prediction result is reduced.

In another embodiment of the present application, refer to fig. 5, which shows a schematic structural diagram of another model detection apparatus 30 provided in the embodiment of the present application. As shown in fig. 5, the model detecting apparatus 30 includes:

the prediction unit 401 is configured to input data to be detected into a target prediction model to obtain a prediction result;

an extracting unit 402, configured to perform feature extraction on data to be detected, and determine a plurality of features corresponding to the data to be detected;

the fusion unit 403 is configured to perform multi-modal fusion processing on the multiple features to obtain target features corresponding to the data to be detected;

a calculating unit 404, configured to determine training sample distribution information corresponding to the target prediction model through the metric model, and perform metric calculation on the training sample distribution information and the target features to obtain a metric result;

a processing unit 405 configured to determine whether the prediction result is correct according to the measurement result.

In some embodiments, the metric model includes an anti-forgetting module and at least one metric computation submodel. Accordingly, the computing unit 404 is specifically configured to obtain training sample distribution information from the anti-forgetting module; calculating target characteristics and training sample distribution information based on at least one measurement calculation sub-model to obtain at least one measurement score information; at least one metric score information is determined as a metric result.

In some embodiments, the processing unit 405 is further configured to output the prediction result if the prediction result is correct; determining the data to be detected as abnormal data under the condition that the prediction result is wrong; storing the abnormal data in a target area; and when the number of the abnormal data in the target area reaches a target threshold value, updating the target prediction model and the measurement model by using all the abnormal data.

In some embodiments, the processing unit 405 is further configured to perform interpolation calculation on the at least one metric score information to obtain an abnormal value of the abnormal data in case of an error in the prediction result; comparing the abnormal value of the abnormal data with an abnormal threshold value to determine the abnormal category of the abnormal data; wherein the exception category includes at least one of: new type samples, difficult to learn samples, or drift samples.

In some embodiments, the processing unit 405 is further configured to determine annotation suggestion information of the abnormal data according to the abnormal category of the abnormal data; sending abnormal data to a user according to the labeling suggestion information of the abnormal data; and receiving the real label of the abnormal data sent by the user.

In some embodiments, the processing unit 405 is further configured to perform statistics on the abnormal data belonging to the new type of sample, the abnormal data belonging to the sample difficult to learn, and the abnormal data belonging to the drift sample, so as to obtain a statistical result; determining a data deviation value according to the statistical result; under the condition that the data deviation value is larger than the deviation threshold value, updating the target prediction model by utilizing all abnormal data and all real labels of the abnormal data based on the first type of updating strategy; under the condition that the data deviation value is smaller than or equal to the deviation threshold value, updating the target prediction model by utilizing all abnormal data and all real labels of the abnormal data based on the second type of updating strategy; and the retention proportion of the original model features of the target prediction model in the first type of updating strategy is smaller than that of the original model features of the target prediction model in the second type of updating strategy.

In some embodiments, the processing unit 405 is further configured to train the original training samples, and build a target prediction model; acquiring target characteristics of an original training sample, and determining training sample distribution information of a target prediction model according to the target characteristics of the original training sample; storing the distribution information of the training samples to an anti-forgetting module; under the condition that the prediction result is correct, updating the distribution information of the training samples in the anti-forgetting module by using the target characteristics of the data to be detected; and under the condition that the prediction result is wrong, determining the target characteristics of the abnormal data to update the distribution information of the training samples in the anti-forgetting module in the process of training the target prediction model by using the abnormal data and the real labels of the abnormal data.

It is understood that in this embodiment, a "unit" may be a part of a circuit, a part of a processor, a part of a program or software, etc., and may also be a module, or may also be non-modular. Moreover, each component in the embodiment may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware or a form of a software functional module.

Based on the understanding that the integrated unit, if implemented in the form of a software functional module and not sold or used as an independent product, may be stored in a computer-readable storage medium, and based on this understanding, a portion that contributes to the technical solution of the present embodiment per se, or all or part of the technical solution may be embodied in the form of a software product stored in a storage medium, and includes several instructions to enable a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute all or part of the steps of the method of the present embodiment. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.

Accordingly, the present embodiment provides a computer storage medium storing a computer program which, when executed by a plurality of processors, implements the steps of the method of any one of the preceding embodiments.

Based on the above-mentioned components of a model detection apparatus 30 and computer storage media, refer to fig. 6, which shows a schematic diagram of a hardware structure of an electronic device 50 according to an embodiment of the present application. As shown in fig. 6, the electronic device 50 may include: a communication interface 501, a memory 502, and a processor 503; the various components are coupled together by a bus device 504. It is understood that bus device 504 is used to enable connected communication between these components. Bus device 504 includes a power bus, a control bus, and a status signal bus in addition to a data bus. But for clarity of illustration the various buses are labeled as bus device 504 in figure 6. The communication interface 501 is used for receiving and sending signals in the process of receiving and sending information with other external network elements;

a memory 502 for storing a computer program capable of running on the processor 503;

a processor 503, configured to execute, when running the computer program:

It will be appreciated that the memory 502 in the embodiments of the subject application can be either volatile memory or nonvolatile memory, or can include both volatile and nonvolatile memory. The non-volatile Memory may be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically Erasable PROM (EEPROM), or a flash Memory. Volatile Memory can be Random Access Memory (RAM), which acts as external cache Memory. By way of example, but not limitation, many forms of RAM are available, such as Static random access memory (Static RAM, SRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic random access memory (Synchronous DRAM, SDRAM), Double Data Rate Synchronous Dynamic random access memory (ddr Data Rate SDRAM, ddr SDRAM), Enhanced Synchronous SDRAM (ESDRAM), Synchronous chained SDRAM (Synchronous link DRAM, SLDRAM), and Direct memory bus RAM (DRRAM). The memory 502 of the apparatus and methods described herein is intended to comprise, without being limited to, these and any other suitable types of memory.

And the processor 503 may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuits of hardware or instructions in the form of software in the processor 503. The Processor 503 may be a general-purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, or discrete hardware components. The various methods, steps, and logic blocks disclosed in the embodiments of the present application may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in the memory 502, and the processor 503 reads the information in the memory 502 and completes the steps of the above method in combination with the hardware thereof.

It is to be understood that the embodiments described herein may be implemented in hardware, software, firmware, middleware, microcode, or any combination thereof. For a hardware implementation, the Processing units may be implemented within one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), general purpose processors, controllers, micro-controllers, microprocessors, other electronic units configured to perform the functions of the present Application, or a combination thereof.

For a software implementation, the techniques of this application may be implemented with modules (e.g., procedures, functions, and so on) that perform the functions of the present application. The software codes may be stored in a memory and executed by a processor. The memory may be implemented within the processor or external to the processor.

Optionally, as another embodiment, the processor 503 is further configured to perform the steps of the method of any one of the preceding embodiments when running the computer program.

In a further embodiment of the present application, based on the composition diagram of the model detection apparatus 30, refer to fig. 7, which shows a composition structure diagram of another electronic device 50 provided in an embodiment of the present application. As shown in fig. 7, the electronic device 50 includes at least the model detection apparatus 30 of any of the previous embodiments.

For the electronic device 50, because it includes the model detection device 30, by performing feature extraction and multi-modal fusion processing on the data to be detected, the obtained target features can more comprehensively reflect the relevant information of the data to be detected, and the accuracy of subsequent judgment is improved; in addition, whether the data to be detected is abnormal is determined by using a measurement result between the target characteristic of the data to be detected and the distribution information of the training sample, so that whether the prediction result of the target prediction model is correct is judged, the target prediction model in a production environment can be supervised, and the loss caused by the wrong prediction result is reduced.

The above description is only a preferred embodiment of the present application, and is not intended to limit the scope of the present application.

It should be noted that, in the present application, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The above-mentioned serial numbers of the embodiments of the present application are merely for description and do not represent the merits of the embodiments.

The methods disclosed in the several method embodiments provided in the present application may be combined arbitrarily without conflict to obtain new method embodiments.

Features disclosed in several of the product embodiments provided in the present application may be combined in any combination to yield new product embodiments without conflict.

The features disclosed in the several method or apparatus embodiments provided in the present application may be combined arbitrarily, without conflict, to arrive at new method embodiments or apparatus embodiments.

The above description is only for the specific embodiments of the present application, but the scope of the present application is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present application, and shall be covered by the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method of model detection, the method comprising:

extracting features of the data to be detected, and determining a plurality of features corresponding to the data to be detected;

determining training sample distribution information corresponding to the target prediction model through a measurement model, and performing measurement calculation on the training sample distribution information and the target characteristics to obtain a measurement result;

2. The method according to claim 1, wherein the metric model includes an anti-forgetting module and at least one metric calculation submodel, and the determining, by the metric model, training sample distribution information corresponding to the target prediction model, and performing metric calculation on the training sample distribution information and the target feature to obtain a metric result includes:

acquiring the training sample distribution information from the anti-forgetting module;

calculating the target characteristics and the training sample distribution information based on the at least one metric calculation submodel to obtain at least one metric score information;

determining the at least one metric score information as the metric result.

3. The method of claim 2, further comprising:

determining the data to be detected as abnormal data under the condition that the prediction result is wrong; storing the abnormal data in a target area; and when the quantity of abnormal data in the target area reaches a target threshold value, updating the target prediction model and the measurement model by using all abnormal data.

4. The method of claim 3, in the event that the prediction is erroneous, the method further comprising:

performing interpolation calculation on the at least one metric score information to obtain an abnormal value of the abnormal data;

comparing an abnormal value of the abnormal data with an abnormal threshold value to determine an abnormal category of the abnormal data; wherein the anomaly category includes at least one of: new type samples, difficult to learn samples, or drift samples.

5. The method of claim 4, after the determining the anomaly category for the anomaly data, the method further comprising:

determining labeling suggestion information of the abnormal data according to the abnormal category of the abnormal data;

sending the abnormal data to a user according to the labeling suggestion information of the abnormal data;

and receiving the real label of the abnormal data sent by the user.

6. The method of claim 5, wherein the updating the target prediction model with all anomaly data comprises:

counting the abnormal data belonging to the new type of samples, the abnormal data belonging to the samples difficult to learn and the abnormal data belonging to the drift samples to obtain a statistical result;

determining a data deviation value according to the statistical result;

under the condition that the data deviation value is larger than a deviation threshold value, updating the target prediction model by utilizing all abnormal data and all real labels of the abnormal data based on a first type of updating strategy;

under the condition that the data deviation value is smaller than or equal to a deviation threshold value, updating the target prediction model by using all abnormal data and all real tags of the abnormal data based on a second type of updating strategy;

wherein the retention ratio of the original model features of the target prediction model in the first class of update strategies is smaller than the retention ratio of the original model features of the target prediction model in the second class of update strategies.

7. The method according to any one of claims 3-6, further comprising:

training an original training sample, and establishing the target prediction model; and the number of the first and second groups,

acquiring target characteristics of the original training sample, and determining training sample distribution information of the target prediction model according to the target characteristics of the original training sample;

storing the training sample distribution information to the anti-forgetting module;

under the condition that the prediction result is correct, updating the training sample distribution information in the anti-forgetting module by using the target characteristics of the data to be detected;

8. A model detection apparatus comprising:

the extraction unit is configured to perform feature extraction on the data to be detected and determine a plurality of features corresponding to the data to be detected;

the fusion unit is configured to perform multi-mode fusion processing on the plurality of features to obtain target features corresponding to the data to be detected;

the calculation unit is configured to determine training sample distribution information corresponding to the target prediction model through a measurement model, and perform measurement calculation on the training sample distribution information and the target characteristics to obtain a measurement result;

9. An electronic device comprising a memory and a processor; wherein the content of the first and second substances,

the memory for storing a computer program operable on the processor;

the processor, when running the computer program, is configured to perform the method of any of claims 1 to 7.

10. A computer storage medium, characterized in that the computer storage medium stores a computer program which, when executed by at least one processor, implements the method of any one of claims 1 to 7.