Autonomous learning method, device, electronic device and machine-readable storage medium

Abstract

This application provides an autonomous learning method, apparatus, electronic device, and machine-readable storage medium. The method comprises: when it is determined that an autonomous learning trigger condition has been met, collecting current training data; online training a first model based on the current training data, historical high-value samples, and acquired labeled data to obtain a second model; and updating the first model based on the second model. This method can achieve continuous model evolution and effectively mitigate the model's forgetting of old knowledge while ensuring the model's performance in processing new data.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to an autonomous learning method, device, electronic device, and machine-readable storage medium. Background Art

[0002] The intelligent analysis and autonomous learning system in video surveillance refers to an intelligent system that can use unlabeled data from users' sites for autonomous learning and continuously improve the generalization capabilities of online model scenarios.

[0003] Traditional autonomous learning systems generally lack the ability to continuously learn. On the one hand, this leads to a waste of hardware training resources on the end-side. On the other hand, they are unable to cope with the continuous changes in online distribution, resulting in gradual performance degradation after a single autonomous learning session. Summary of the Invention

[0004] In view of this, the present application provides an autonomous learning method, device, electronic device and machine-readable storage medium to cope with the continuous changes in online data distribution and optimize model performance.

[0005] Specifically, this application is implemented through the following technical solutions:

[0006] According to a first aspect of an embodiment of the present application, there is provided an autonomous learning method, comprising:

[0007] When it is determined that the autonomous learning trigger condition is met, current training data is collected; the current training data includes unlabeled data generated within a first preset time period before the target time point, and / or unlabeled data generated within a second preset time period after the target time point; the target time point is the time point when the autonomous learning trigger condition is determined to be met;

[0008] Based on the current training data, historical high-value samples, and the acquired labeled data, the first model is trained online to obtain a second model; wherein the first model is the currently used model, and the historical high-value samples are samples in the historical autonomous learning process of the currently used model, in which the difference in output results of the model before and after autonomous learning is greater than a preset threshold;

[0009] The first model is updated according to the second model.

[0010] According to a second aspect of an embodiment of the present application, there is provided an autonomous learning device, comprising:

[0011] A determination unit, used to determine whether the autonomous learning triggering condition is met;

[0012] a data collection unit configured to collect current training data when the determining unit determines that the autonomous learning trigger condition is satisfied; the current training data includes unlabeled data generated within a first preset time period before a target time point, and / or unlabeled data generated within a second preset time period after the target time point; the target time point is the time point at which the autonomous learning trigger condition is determined to be satisfied;

[0013] a training unit for performing online training on the first model based on the current training data, the historical high-value samples, and the acquired labeled data to obtain a second model; wherein the first model is the currently used model, and the historical high-value samples are samples in the historical autonomous learning process of the currently used model, for which the difference in output results between the model before and after autonomous learning is greater than a preset threshold;

[0014] An updating unit, configured to update the first model according to the second model.

[0015] According to a third aspect of an embodiment of the present application, an electronic device is provided, comprising a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor is used to execute the machine-executable instructions to implement the method provided in the first aspect.

[0016] According to a fourth aspect of an embodiment of the present application, a machine-readable storage medium is provided, wherein the machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a processor, the method provided in the first aspect is implemented.

[0017] The technical solution provided by this application can at least bring the following beneficial effects:

[0018] By collecting unlabeled data generated within a first preset time period before the current time point and / or unlabeled data generated within a second preset time period after the current time point when determining that the trigger conditions for autonomous learning are met, and using the collected unlabeled data as current training data, the first model is trained online based on the current training data, historical high-value samples, and the acquired labeled data, and the first model is updated based on the trained second model, thereby achieving continuous evolution of the model and effectively alleviating the model's forgetting of old knowledge while ensuring the model's processing performance for new data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of an autonomous learning method shown in an exemplary embodiment of the present application;

[0020] Figure 2 This is a complete flowchart of autonomous learning shown in an exemplary embodiment of the present application;

[0021] Figure 3 is a structural diagram of an autonomous learning device shown in an exemplary embodiment of the present application;

[0022] Figure 4 It is a schematic diagram of the hardware structure of an electronic device shown in an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0023] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0024] The terms used in this application are for the purpose of describing particular embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0025] In order to enable those skilled in the art to better understand the technical solutions provided by the embodiments of the present application, and to make the above-mentioned purposes, features and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are further described in detail below with reference to the accompanying drawings.

[0026] See Figure 1 , is a flow chart of an autonomous learning method provided in an embodiment of the present application, such as Figure 1 As shown, the autonomous learning method may include the following steps:

[0027] It should be noted that the serial numbers of the steps in the embodiments of the present application do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0028] Step S100: When it is determined that the autonomous learning trigger condition is met, current training data is collected; the current training data includes unlabeled data generated within a first preset time period before the target time point, and / or unlabeled data generated within a second preset time period after the target time point; the target time point is the time point at which the autonomous learning trigger condition is determined to be met.

[0029] In an embodiment of the present application, in order to improve the generalization ability of online model scenarios and cope with the continuous changes in the distribution of online data, when it is determined that the autonomous learning trigger conditions are met, unlabeled data generated within a preset time period (referred to as the first preset time period herein) before the current time (i.e., the time point at which the autonomous learning trigger conditions are determined to be met, referred to as the target time point herein) and / or unlabeled data generated within a preset time period (referred to as the second preset time period herein) after the target time point can be collected (referred to as current training data herein) for autonomous learning training of the currently used model (referred to as the first model herein).

[0030] Exemplarily, the current training data may be data processed by the first model, which may include alarm data (ie, data that triggers an alarm) and non-alarm data (ie, data that does not trigger an alarm).

[0031] Illustratively, the first preset duration and the second preset duration may be the same or different.

[0032] Exemplarily, one of the first preset time length and the second preset time length can be 0, that is, only the unlabeled data before the target time point can be collected as the current training data, or only the unlabeled data after the target time point can be collected as the current training data.

[0033] Exemplarily, when the current training data includes unlabeled data generated within a second preset time period after the target time point, the model update can be started when the current training data collection is completed.

[0034] Step S110: Based on the current training data, historical high-value samples, and the acquired labeled data, the first model is trained online to obtain the second model.

[0035] In an embodiment of the present application, in order to ensure the accuracy of the model obtained by autonomous learning and training, in the process of autonomous learning and training of the first model, in addition to using the current training data obtained in step S100, some labeled data can also be used.

[0036] In addition, in order to ensure the processing performance of the autonomous learning and updating model for new data and alleviate the model's forgetting of old knowledge, when the model is autonomously learned and updated, the training data can include historical training data in addition to the above-mentioned current training data and labeled data.

[0037] Considering that too much historical training data will increase the amount of training data and reduce learning efficiency, and too little historical training data may not be able to guarantee the effect of alleviating the forgetting of old knowledge, we can avoid the impact of too much historical training data on learning efficiency and at the same time, avoid forgetting old knowledge by selecting high-value historical data (referred to as high-value historical samples in this article) to participate in the autonomous learning and training of the model.

[0038] Exemplarily, a high-value sample may be a sample in which the difference in output results of the model before and after autonomous learning is greater than a preset threshold during the historical autonomous learning process of the currently used model.

[0039] For example, in each process of autonomous learning training of the model according to the method described in step S100 to step S120, assuming that the current autonomous learning is the tth (t≥1)th autonomous learning, the model before autonomous learning is recorded as M t , the model after autonomous learning is recorded as M t+1 , then high-value historical samples can be included in the model M t The output result on M t+1 The difference between the output results is greater than the preset threshold.

[0040] For example, high-value historical samples can be selected from the current training data. Each time the model is trained autonomously in the manner described in steps S100 to S120, the model M can be selected based on the values ​​of the samples in the current training data before training. t (ie, the first model), and the trained model M t+1 (i.e., the second model), select high-value historical samples from the current training data, and add the selected high-value historical samples to the high-value sample library.

[0041] For example, when the tth autonomous learning is completed, samples whose output results on model Mt and the output results on Mt+1 are different from each other by more than a preset threshold can be obtained from the current training data collected in the tth autonomous learning as high-value historical samples.

[0042] It should be noted that when the model is first autonomously learned, historical high-value samples may not be used, or a certain number of high-value samples may be selected as historical high-value samples through manual selection. This embodiment of the present application does not limit this.

[0043] Accordingly, the first model can be trained online using a semi-supervised learning method based on the current training data, historical high-value training samples, and the acquired labeled data to obtain a trained model (referred to as the second model in this article).

[0044] In one example, considering that labeled data affects the efficiency of model learning, when there is a lot of labeled data, the model learning time will be longer.

[0045] In addition, considering that the computing power of front-end devices is usually weak, if autonomous learning is applied to front-end devices, the increase in the amount of labeled data will significantly reduce the efficiency of autonomous learning of front-end devices.

[0046] Therefore, in order to improve the efficiency of model learning while ensuring the accuracy of model training, in the process of autonomous learning in the above manner, the amount of labeled data can be smaller than the amount of current training data.

[0047] For example, the ratio of the amount of labeled data to the amount of current training data may be less than a preset ratio threshold.

[0048] For example, labeled data can be obtained through manual labeling or model-assisted labeling.

[0049] Step S120: Update the first model according to the second model.

[0050] In the embodiment of the present application, when the second model is obtained by training according to the method described in step S110, the first model can be updated according to the second model.

[0051] It should be noted that in the embodiment of the present application, the autonomous learning of the model is a continuous process. After a model update is performed in the manner described in steps S100 to S120, if it is determined again that the autonomous learning trigger conditions are met, the second model obtained last time can be used as the new first model, and the model update can be performed again in accordance with steps S100 to S120.

[0052] For example, in order to reduce the difficulty of data acquisition and improve the efficiency of data acquisition, the model is updated each time in the manner described in steps S100 to S120 , and the labeled data used may be the same.

[0053] It can be seen that in Figure 1 In the method flow shown, by determining that the autonomous learning trigger conditions are met, collecting unlabeled data generated within a first preset time period before the current time point, and / or unlabeled data generated within a second preset time period after the current time point, and using the collected unlabeled data as current training data, the first model is trained online based on the current training data, historical high-value samples, and the acquired labeled data, and the first model is updated based on the trained second model, thereby achieving continuous evolution of the model and effectively alleviating the model's forgetting of old knowledge while ensuring the model's processing performance for new data.

[0054] In some embodiments, in step S100, satisfying the autonomous learning triggering condition may include:

[0055] The accuracy of the online inference result of the first model is lower than a preset accuracy threshold; and / or,

[0056] Detecting an autonomous learning trigger instruction; and / or,

[0057] Reach the preset independent learning time point.

[0058] Illustratively, autonomous learning may be triggered automatically and / or manually.

[0059] For example, considering that after a model is updated through autonomous learning, the performance of the updated model will gradually degrade with the emergence of new data, and thus the accuracy of the model's online reasoning results will also gradually decrease.

[0060] Accordingly, in order to ensure model performance, when the accuracy of the online inference results of the first model is lower than the preset accuracy threshold, it can be determined that the autonomous learning trigger condition is met, and the model can be updated through autonomous learning in the manner described in steps S100 to S120.

[0061] Exemplarily, in an embodiment of the present application, the autonomous learning update of the model can be triggered periodically, or the autonomous learning update of the model can be triggered manually.

[0062] Accordingly, when an autonomous learning trigger instruction is detected or a preset autonomous learning time point is reached, it can be determined that the autonomous learning trigger condition is met.

[0063] In some embodiments, high-value historical samples in the high-value historical sample library for online training are screened in the following manner:

[0064] When the t-th autonomous learning is completed, obtaining a first sample in which the difference between the output result on the model Mt and the output result on Mt+1 is greater than a preset threshold;

[0065] The first sample is de-similarized using a clustering method to obtain a second sample, which is then added to a high-value historical sample library.

[0066] For example, considering that some samples with high similarity may be collected during the current training data collection process, and these samples with high similarity may all meet the conditions of high-value historical samples, resulting in the presence of samples with high similarity in high-value historical samples, making the types of samples in high-value historical samples fewer, which is not conducive to improving model performance.

[0067] For example, taking the perimeter control scenario as an example, a monitoring front-end device is set up in the wall area of ​​the area, and the image data collected by the monitoring front-end device is used to determine whether someone has climbed over the wall. In this scenario, if there are large trees near the wall, the leaves swaying above the wall due to the wind may cause frequent false alarms. Although the value of this data is high, the similarity is also high.

[0068] Accordingly, for samples (referred to as first samples in this article) whose output results, screened out based on the model before and after training, differ by more than a preset threshold, the first samples can be clustered using a clustering method to obtain at least one cluster; for any cluster, the support set of the cluster can be determined based on the first samples in the cluster; wherein the distance (such as Euclidean distance) between the first samples in the same support set exceeds the preset threshold, and the samples in the support set can effectively characterize the distribution of the first samples in the cluster.

[0069] For any cluster, the first sample in the support set of the cluster can be determined as a high-value sample in the cluster (which can be called a second sample), and the remaining samples are removed as similar samples to the second sample.

[0070] When the second samples in each cluster are obtained in the above manner, the determined second samples can be added to the high-value historical sample library as high-value historical samples.

[0071] In one example, adding the second sample to the high-value historical sample library may include:

[0072] For any second sample, when the number of high-value historical samples in the high-value historical sample library reaches a preset maximum number before the second sample is added to the high-value historical sample library, the high-value historical sample with the earliest addition time in the high-value historical sample library is deleted, and the second sample is added to the high-value historical sample library.

[0073] For example, the samples in the high-value historical sample library can be updated in a first-in-first-out manner with reference to the form of a queue. When the samples in the high-value historical sample library reach the maximum storage capacity, the samples can be deleted in order of storage time from the earliest to the latest.

[0074] Accordingly, for any second sample determined in the above manner, before adding the second sample to the high-value historical sample library, it can be determined whether the number of high-value historical samples in the current high-value historical sample library reaches a preset maximum number.

[0075] When the number of high-value historical samples in the current high-value historical sample library reaches a preset maximum number, the high-value historical sample with the earliest addition time in the high-value historical sample library can be deleted, and the second sample can be added to the high-value historical sample library.

[0076] In some embodiments, in step S120, updating the first model according to the second model may include:

[0077] The second model is quantized online, and the quantized second model is used to replace the first model.

[0078] For example, in order to reduce the computational complexity of the model and improve the processing efficiency of the model, when the first model is autonomously trained to obtain the second model in the above-mentioned manner, the second model can be quantized online, and the quantized second model can be used to replace the first model, thereby realizing automatic evolution and upgrading of the model without human intervention.

[0079] In order to enable those skilled in the art to better understand the technical solutions provided by the embodiments of the present application, the technical solutions provided by the embodiments of the present application are described below with reference to specific examples.

[0080] In an embodiment of the present application, an autonomous learning method that supports continuous evolution is provided. Through a reasonable data screening strategy, high-value historical samples that have appeared in the current scene are continuously accumulated, and in the subsequent autonomous learning process, high-value historical samples and current samples (i.e., the above-mentioned current training data) are used for training at the same time, thereby ensuring that the trained model can solve the problems of false positives / missing negatives that have recently appeared, and at the same time solve the problems that have appeared in history, avoiding the phenomenon of knowledge forgetting, thereby ensuring that the autonomous learning system has the ability to continuously evolve while not forgetting the knowledge that has been learned.

[0081] In this embodiment, the autonomous learning solution can be divided into five stages during implementation: triggered learning, current sample collection, online training, high-value historical sample screening, and model launch.

[0082] 1. Triggering the Learning Phase

[0083] For example, during the triggered learning phase, online reasoning and diagnostics can be performed on the video stream data and model results to determine if there are any anomalies in the alarm data, thereby determining whether autonomous learning should be triggered. Alternatively, a manual trigger signal or a timed trigger can be introduced to trigger autonomous learning.

[0084] When it is determined that autonomous learning needs to be triggered (ie, it is determined that the autonomous learning triggering conditions are met), the current sample collection phase can be entered.

[0085] 2. Current Sample Collection Phase

[0086] For example, during the current sample collection phase, T days of data before and after the current time point (the time point at which autonomous learning is required) may be collected for autonomous learning training.

[0087] For example, the current sample may include alarm data and non-alarm data (the images are stored in the historical records and are marked as to whether an alarm is triggered, but they are all unlabeled data).

[0088] 3. Online training phase

[0089] For example, during the online training phase, current samples (data obtained during the current sample collection phase), high-value historical sample data, and a small amount of labeled data are input, and online training is performed using a semi-supervised learning method to improve recall and suppress false positives.

[0090] Exemplarily, online training methods may include but are not limited to Self-Training / Domain Adaptation algorithms.

[0091] For example, after each online training, high-value historical samples can be mined and stored in a high-value historical sample library. Therefore, during the first training, the high-value historical sample library can be empty, and only current samples and a small amount of labeled data need to be used for training.

[0092] 4. High-value historical sample screening stage

[0093] For example, in the high-value historical sample screening stage, the pre-training model M can be mined t And the trained model M t+1 Differences at the sample level (the greater the difference in the results obtained by the two models for the same sample, the higher the value).

[0094] Exemplarily, if the difference in output results of the same sample on different models is greater than a preset threshold, the sample is determined to be a high-value sample.

[0095] For example, let the sample be x and the model output be M(x). Then, the samples whose output difference between the pre-training model and the post-training model is greater than a preset threshold can be considered as high-value historical samples:

[0096] x:|M t +1(x)-M t (x)|>τ

[0097] For example, clustering or other methods can be used to find support sets in difference samples, remove samples with high similarity, screen out a certain number of high-value historical samples, store them in a high-value historical sample library, and use them for the next autonomous learning training.

[0098] For example, a high-value historical sample library can be used as a knowledge base, retaining high-value knowledge points that have appeared in the scenario. Using this sample library for learning can effectively alleviate the problem of knowledge forgetting and obtain continuous evolution capabilities.

[0099] Since high-value historical samples are screened during each training session and added to the high-value historical sample library, the number of samples in the high-value historical sample library will continue to grow after multiple training sessions. Therefore, the high-value historical sample library can be maintained and deleted.

[0100] For example, during implementation, the high-value historical sample library can be updated in a queue manner: that is, a size S is set, and after exceeding S, the first-in-first-out rule is used to replace the earliest high-value historical samples with the latest high-value historical samples to avoid continuous expansion of the high-value historical sample library.

[0101] 5. Model launch stage

[0102] For example, after model training is completed, the model can be quantized online to reduce the amount of model calculation, and then deployed online to replace the old model, thereby achieving automatic evolution and upgrading of the model without human intervention.

[0103] For example, the complete process of autonomous learning can be as follows: Figure 2 shown.

[0104] The above describes the method provided by this application. The following describes the device provided by this application:

[0105] See Figure 3 , is a structural diagram of an autonomous learning device provided in an embodiment of the present application, such as Figure 3 As shown, the autonomous learning device may include:

[0106] A determination unit 310 is configured to determine whether an autonomous learning triggering condition is met;

[0107] The data collection unit 320 is configured to collect current training data when the determination unit 310 determines that the autonomous learning trigger condition is satisfied; the current training data includes unlabeled data generated within a first preset time period before a target time point, and / or unlabeled data generated within a second preset time period after the target time point; the target time point is the time point at which the autonomous learning trigger condition is determined to be satisfied;

[0108] A training unit 330 is configured to perform online training on the first model based on the current training data, the historical high-value samples, and the acquired labeled data to obtain a second model; wherein the first model is the currently used model, and the historical high-value samples are samples in the historical autonomous learning process of the currently used model, for which the difference in output results between the model before and after autonomous learning is greater than a preset threshold;

[0109] The updating unit 340 is configured to update the first model according to the second model.

[0110] In some embodiments, satisfying the autonomous learning triggering condition includes:

[0111] The accuracy of the online inference result of the first model is lower than a preset accuracy threshold; and / or,

[0112] Detecting an autonomous learning trigger instruction; and / or,

[0113] Reach the preset independent learning time point.

[0114] In some embodiments, the data collection unit 320 selects high-value historical samples for online training in the following manner:

[0115] When the t-th autonomous learning is completed, obtaining the first sample where the difference between the output result on the model Mt and the output result on Mt+1 is greater than a preset threshold; Mt is the model before the t-th autonomous learning, Mt+1 is the model after the t-th autonomous learning, and t≥1;

[0116] The first sample is subjected to de-similarity processing by using a clustering method to obtain a second sample, and the second sample is added to the high-value historical sample library.

[0117] In some embodiments, the data collection unit 320 adds the second sample to the high-value historical sample library, including:

[0118] For any second sample, when the number of high-value historical samples in the high-value historical sample library reaches a preset maximum number before the second sample is added to the high-value historical sample library, the high-value historical sample with the earliest addition time in the high-value historical sample library is deleted, and the second sample is added to the high-value historical sample library.

[0119] In some embodiments, the updating unit 340 updates the first model according to the second model, including:

[0120] The second model is quantized online, and the first model is replaced by the quantized second model.

[0121] An embodiment of the present application provides an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor is used to execute the machine-executable instructions to implement the autonomous learning method described above.

[0122] See Figure 4 , is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application. The electronic device may include a processor 401 and a memory 402 storing machine-executable instructions. The processor 401 and the memory 402 may communicate via a system bus 403. Furthermore, by reading and executing the machine-executable instructions corresponding to the autonomous learning logic in the memory 402, the processor 401 may perform the autonomous learning method described above.

[0123] The memory 402 mentioned herein may be any electronic, magnetic, optical, or other physical storage device that may contain or store information, such as executable instructions, data, and the like. For example, the machine-readable storage medium may be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, a storage drive (such as a hard disk drive), a solid-state drive, any type of storage disk (such as a CD, DVD, etc.), or similar storage media, or a combination thereof.

[0124] In some embodiments, a machine-readable storage medium is also provided. Figure 4 The memory 402 in the machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are executed by the processor, the autonomous learning method described above is implemented. For example, the storage medium may be a ROM, RAM, CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0125] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0126] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. An autonomous learning method, characterized in that: include: When it is determined that the autonomous learning triggering conditions are met, the current training data is collected; The current training data includes unlabeled data generated within a first preset time period before the target time point, and / or unlabeled data generated within a second preset time period after the target time point; The target time point is a time point at which the autonomous learning trigger condition is determined to be satisfied; wherein, the determination that the autonomous learning trigger condition is satisfied includes: the accuracy of the result of online inference of the video stream data using the first model is lower than a preset accuracy threshold; Based on the current training data, historical high-value samples, and the acquired labeled data, the first model is trained online to obtain a second model; wherein the first model is the currently used model, and the historical high-value samples are samples in the historical autonomous learning process of the currently used model, in which the difference in output results of the model before and after autonomous learning is greater than a preset threshold; updating the first model according to the second model; Among them, the high-value historical samples in the high-value historical sample library used for online training are screened in the following way: When the tth autonomous learning is completed, the t The output result on M t+1 The difference between the output results on the first sample is greater than the preset threshold; the M t is the model before the t-th autonomous learning, the M t+1 is the model after the t-th autonomous learning, t≥1; wherein the first sample is image data collected by the monitoring front-end device; The first sample is subjected to de-similarity processing by using a clustering method to obtain a second sample, and the second sample is added to the high-value historical sample library.

2. The method according to claim 1, characterized in that The autonomous learning triggering condition is satisfied, and further includes: Detecting an autonomous learning trigger instruction; and / or, Reach the preset independent learning time point.

3. The method according to claim 1, characterized in that Adding the second sample to the high-value historical sample library includes: For any second sample, when the number of high-value historical samples in the high-value historical sample library reaches a preset maximum number before the second sample is added to the high-value historical sample library, the high-value historical sample with the earliest addition time in the high-value historical sample library is deleted, and the second sample is added to the high-value historical sample library.

4. The method according to claim 1, wherein The updating of the first model according to the second model includes: The second model is quantized online, and the first model is replaced by the quantized second model.

5. An autonomous learning device, characterized in that: include: A determination unit, used to determine whether the autonomous learning triggering condition is met; a data collection unit configured to collect current training data when the determining unit determines that the autonomous learning trigger condition is satisfied; the current training data comprising unlabeled data generated within a first preset time period before the target time point, and / or unlabeled data generated within a second preset time period after the target time point; The target time point is a time point at which the autonomous learning trigger condition is determined to be satisfied; wherein, the determination that the autonomous learning trigger condition is satisfied includes: the accuracy of the result of online inference of the video stream data using the first model is lower than a preset accuracy threshold; a training unit for performing online training on the first model based on the current training data, the historical high-value samples, and the acquired labeled data to obtain a second model; wherein the first model is the currently used model, and the historical high-value samples are samples in the historical autonomous learning process of the currently used model, for which the difference in output results between the model before and after autonomous learning is greater than a preset threshold; an updating unit, configured to update the first model according to the second model; The data collection unit selects high-value historical samples for online training in the following manner: Upon completion of the t-th autonomous learning, a first sample is obtained in which the difference between the output result on the model Mt and the output result on Mt+1 is greater than a preset threshold; Mt is the model before the t-th autonomous learning, Mt+1 is the model after the t-th autonomous learning, and t≥1; wherein the first sample is image data collected by the monitoring front-end device; The first sample is subjected to de-similarity processing by using a clustering method to obtain a second sample, and the second sample is added to the high-value historical sample library.

6. The device according to claim 5, characterized in that The autonomous learning triggering condition is satisfied, and further includes: Detecting an autonomous learning trigger instruction; and / or, Reach the preset independent learning time point.

7. The device according to claim 5, characterized in that The data collection unit adds the second sample to the high-value historical sample library, including: For any second sample, when the number of high-value historical samples in the high-value historical sample library reaches a preset maximum number before the second sample is added to the high-value historical sample library, the high-value historical sample with the earliest addition time in the high-value historical sample library is deleted, and the second sample is added to the high-value historical sample library.

8. The device according to claim 5, characterized in that The updating unit updates the first model according to the second model, including: The second model is quantized online, and the first model is replaced by the quantized second model.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor is configured to execute the machine-executable instructions to implement the method according to any one of claims 1 to 4.

10. A machine-readable storage medium, characterized in that The machine-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images