Self-learning method, system and device for updating algorithm model and storage medium

Abstract

This application provides a self-learning method, system, device, and storage medium for updating algorithm models. In this application embodiment, a self-learning closed-loop chain is formed, consisting of data collection based on tracking points, sample construction based on active learning, and fine-tuning training based on incremental learning. This supports automated iteration of algorithm model updates for various model paradigms, reduces the manual maintenance cost of algorithm model updates, and improves the update efficiency and intelligence of algorithm models.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a self-learning method, system, device and storage medium for updating algorithm models. Background Technology

[0002] Cloud service providers offer intelligent industry solutions to customers through algorithmic models. These models include, but are not limited to, models with traffic light optimization capabilities, models with real-time travel data prediction capabilities, and models with intelligent recommendation capabilities. In practical applications, before the algorithmic model is deployed online, it undergoes fine-tuning training in an offline environment using the customer's application data (also known as offline fine-tuning) to ensure good model performance after deployment. Even after deployment, the algorithmic model continues to be iterated and updated to maintain optimal performance.

[0003] Currently, the most common way to update an algorithm model after it has been deployed online is to manually collect feedback data from the algorithm model, manually construct a sample dataset, and retrain the algorithm model based on the sample dataset to obtain the updated algorithm model. This update method has high manpower costs and low update efficiency and intelligence. Summary of the Invention

[0004] This application provides a self-learning method, system, device, and storage medium for updating algorithm models, thereby improving the updating efficiency and intelligence of algorithm models.

[0005] This application provides a self-learning method for updating an algorithm model, comprising: in response to a model update event being triggered, acquiring self-learning task configuration information, the self-learning task configuration information including basic information, sample construction configuration information, and fine-tuning training configuration information, the basic information including a first acquisition method for acquiring tracking feedback data, the tracking feedback data referring to relevant information that provides feedback on the online running effect of the algorithm model; acquiring tracking feedback data according to the first acquisition method; constructing samples based on an active learning method according to the tracking feedback data and the sample construction configuration information to obtain the current training sample dataset required for this model update; and fine-tuning the algorithm model before this update based on an incremental learning method according to the fine-tuning training configuration information and the current training sample dataset to obtain the updated algorithm model.

[0006] This application embodiment also provides a self-learning system for updating algorithm models. The system is deployed in a cloud server and includes: a control module, and a human-computer interaction module, a sample construction module, a model training module, and a model evaluation module, all connected to the control module. The human-computer interaction module is used to display a self-learning task configuration interface in response to a user-inputted task configuration trigger operation. In response to a configuration operation triggered by the user through the self-learning task configuration interface, it configures self-learning task configuration information and sends the self-learning task configuration information to the control module. The self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information, and model evaluation configuration information. The basic information includes a first acquisition method for obtaining event tracking feedback data and a second acquisition method for obtaining historical test sample datasets. The model evaluation configuration information includes target performance test metrics and target performance evaluation algorithms. The sample construction module is used to obtain event tracking feedback data according to the first acquisition method under the control of the control module, and according to... The system uses the tracking feedback data and sample construction configuration information to construct samples based on active learning, obtaining the training sample dataset and test sample dataset required for this model update. The model training module, under the control of the control module, performs incremental learning-based fine-tuning training on the algorithm model before the update, based on the fine-tuning training configuration information and the current training sample dataset, to obtain the updated algorithm model. The model evaluation module, under the control of the control module, obtains the historical test sample dataset using a second acquisition method. It then uses the historical test sample dataset and the current test sample dataset to perform performance tests on the algorithm model before and after the update, targeting specific performance metrics. Finally, it uses a target performance evaluation algorithm to compare the performance test results of the algorithm model before and after the update, obtaining a comparative evaluation result. Based on the comparative evaluation result, it determines whether the updated algorithm model meets the conditions for online operation.

[0007] This application also provides an electronic device, including: a memory and a processor; the memory for storing a computer program; and the processor coupled to the memory for executing the computer program to perform steps in a self-learning method for updating an algorithm model.

[0008] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement steps in a self-learning method for updating an algorithm model.

[0009] Based on the above, embodiments of this application provide a self-learning method, system, device, and storage medium for updating algorithm models. In these embodiments, a self-learning closed-loop chain is formed, comprising data collection based on event tracking, sample construction based on active learning, and fine-tuning training based on incremental learning. This supports automated iteration of algorithm model updates for various model paradigms, reducing the manual maintenance costs of algorithm model updates and improving the update efficiency and intelligence of the algorithm model. Furthermore, sample construction based on active learning can better select representative, high-quality samples, effectively reducing the amount of sample data, decreasing the demand for computing resources, and enhancing the feasibility of deploying the algorithm model in low-resource environments. Fine-tuning training based on incremental learning avoids catastrophic forgetting of the model, ensuring that the algorithm model can continuously learn knowledge from new data while maintaining its performance on historical data. This enhances the algorithm model's ability to learn from changes in customer application data and its generalization performance, while reducing the consumption of computing resources. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 An exemplary application scenario diagram of model self-learning provided for embodiments of this application; Figure 2 Example configuration interface for self-learning tasks; Figure 3 A flowchart illustrating a self-learning method for updating an algorithm model, provided in an embodiment of this application; Figure 4 A flowchart illustrating another self-learning method for updating an algorithm model provided in this application embodiment; Figure 5 A flowchart illustrating another self-learning method for updating an algorithm model provided in this application embodiment; Figure 6 Examples of relationships between algorithm models; Figure 7 This is a schematic diagram of the structure of a self-learning system for updating algorithm models provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the access relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone, where A and B can be singular or plural. In the textual description of this application, the character " / " generally indicates that the preceding and following associated objects have an "or" relationship. Furthermore, in the embodiments of this application, "first," "second," "third," etc., are only used to distinguish the content of different objects and have no other special meaning.

[0013] The following is a description of some terms used in the embodiments of this application: Active Learning is a machine learning or artificial intelligence method that actively selects and labels the most valuable samples. Its goal is to achieve good model performance using as few high-quality labeled samples as possible. The general idea behind Active Learning is to use machine learning methods to select samples that are relatively "difficult" to classify and have high uncertainty, then have them manually reviewed and approved. The manually labeled data is then used to train a supervised or semi-supervised learning model, gradually improving the model's performance and integrating human experience into the machine learning model. For more information on Active Learning, please refer to related technical articles.

[0014] Incremental learning ensures that the algorithm model continuously learns new knowledge under limited computational and storage resources without degrading generalization performance. Incremental learning has the following characteristics: ① It retains most of the previously learned knowledge while learning new knowledge, meaning the algorithm model performs well on both old and new tasks. ② Computational power and memory should increase steadily or slowly with the number of classes. Ideally, once a task is completed, all observations for that task are discarded. ③ The model can continuously learn new knowledge from new tasks and new data; it is trainable whenever a new task appears at different times. Currently, mainstream incremental learning methods include, but are not limited to: regularization-based incremental learning, replay-based incremental learning, and parameter isolation-based incremental learning. Regularization-based incremental learning corrects gradients by introducing additional loss (such as distillation loss) to protect the model's learned old knowledge. Replay-based incremental learning retains a portion of representative old data to help the model review previously learned knowledge when training on new tasks. Incremental learning based on parameter isolation involves scaling up an older model on a new task while isolating the parameters of the new and old models to varying degrees, thus reducing catastrophic forgetting. For more information on incremental learning, please refer to the related technologies section.

[0015] Fine-tuning training refers to training an already trained algorithm model with a small amount of new training sample dataset to achieve better model performance. For more information on fine-tuning training, please refer to the related technical documentation.

[0016] Currently, the most common method for updating algorithm models after they have been deployed online is to manually collect feedback data from the algorithm model, manually construct a sample dataset, and retrain the algorithm model based on the sample dataset to obtain the updated algorithm model. This method involves high manpower costs and has low update efficiency and a low level of automation.

[0017] In particular, for many situations where customers retrain algorithm models based on sample datasets using their computing resources, the entire algorithm model update process consumes a significant amount of the customer's computing resources, resulting in low update efficiency and a low level of intelligence. Furthermore, customers often face a shortage of computing resources on-site. Due to cost considerations, not only are GPU (Graphics Processing Unit) resources lacking, but even CPU (Central Processing Unit) and memory resources available for model training are very limited. This leads to a significant increase in training time. Especially as mainstream algorithm models strive for higher algorithm performance and generalization capabilities, the number of parameters and the amount of training data required are constantly increasing. How to train models in low-resource environments is a huge challenge.

[0018] Based on the above, this application also provides a self-learning method, system, device, and storage medium for updating algorithm models. In this application embodiment, a self-learning closed-loop chain is formed, consisting of data collection based on tracking points, sample construction based on active learning, and fine-tuning training based on incremental learning. This supports automated iteration of algorithm model updates for various model paradigms, reducing the manual maintenance cost of algorithm model updates and improving the update efficiency and intelligence of the algorithm model. Furthermore, sample construction based on active learning can better select representative high-quality samples, effectively reducing the amount of sample data, reducing the requirements for computing resources, and enhancing the feasibility of the algorithm model in low-resource environments. Fine-tuning training based on incremental learning avoids catastrophic forgetting of the model, ensuring that the algorithm model can continuously learn knowledge from new data while maintaining its performance on historical data, enhancing the algorithm model's learning ability and generalization performance in response to changes in customer application data.

[0019] Figure 1 This diagram illustrates an exemplary application scenario of model self-learning, provided as an embodiment of this application. See also... Figure 1 First, we collect customer application data and perform offline training in an offline environment based on this data to obtain an algorithm model with specific functions. Before releasing the algorithm model online, we can also use customer application data to fine-tune the training in an offline environment, and then release the finely tuned algorithm model online. The algorithm model after being released online is referred to as the online algorithm model.

[0020] Next, after the algorithm model is launched and running online, feedback data (referred to as event tracking data) can be collected using event tracking techniques. Event tracking feedback data refers to relevant information that provides feedback on the effectiveness of the online algorithm model, including but not limited to: model input data, model output results, and annotation results. The model output results (also called model prediction results) refer to the actual output of the algorithm model after the input data is processed. Annotation results refer to the results obtained by labeling the model input data; these can also be called expected output results or actual processing results. Expected output results are the expected output of the algorithm model after the input data is processed. Of course, the information to be collected for event tracking feedback data can be defined as needed based on the model paradigm of the algorithm model, without any restrictions. Model paradigms include, but are not limited to: text matching model paradigm, classification model paradigm, generative model paradigm, entity extraction model paradigm, etc.

[0021] Next, we move on to the sample construction stage. In this stage, the event tracking feedback data is filtered based on an active learning approach, and the filtered event tracking feedback data is used to construct training sample datasets and test sample datasets. Next, we move on to the fine-tuning training phase. In this phase, the current online algorithm model is fine-tuned using an incremental learning approach using the training sample dataset to obtain an updated algorithm model.

[0022] Next, we proceed to the model evaluation stage. In this stage, we use the test sample dataset to perform performance tests on the algorithm models before and after the update, and then compare the performance test results of the algorithm models before and after the update to obtain a comparative evaluation result of the algorithm model before and after the update.

[0023] Next, the deployment phase begins. In this phase, if the evaluation results indicate that the updated algorithm model performs better than the original online algorithm model, then the updated algorithm model is ready for deployment. At this point, the updated algorithm model is deployed to the online environment to complete the model update task. If the evaluation results indicate that the updated algorithm model performs worse than the original online algorithm model, then the updated algorithm model is not ready for deployment. In this case, the algorithm model in the online environment remains the original algorithm model.

[0024] Based on the above, after the algorithm model is launched, a closed-loop chain for model updates is formed by collecting data from tracking points, constructing high-quality samples based on active learning, fine-tuning training using incremental learning, comparing and evaluating the algorithm model before and after the update, and automatically deciding whether to release an upgrade. This automatically completes the self-learning task of updating the algorithm model. This self-learning task can be executed by the cloud providing cloud computing resources, without relying on the customer's computing resources, improving the efficiency and intelligence of algorithm model updates and enhancing the user experience. Furthermore, even when using the customer's computing resources for model updates, compared to existing methods, it significantly reduces the consumption of the customer's computing resources and improves the efficiency and intelligence of model updates using customer computing resources.

[0025] certainly, Figure 1 The application scenario shown is merely an exemplary one, and the embodiments of this application do not limit the application scenario.

[0026] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0027] In this embodiment, the iterative updates of the algorithm model can be automatically managed by creating self-learning tasks. In practical applications, users can flexibly configure self-learning tasks as needed. For example, a configured self-learning task can perform iterative training on the algorithm model. Another example is that a configured self-learning task can perform iterative training and performance testing on the algorithm model. Yet another example is that a configured self-learning task can perform iterative training, performance testing, comparative evaluation, and determine whether to release the algorithm model online; there are no limitations on this.

[0028] In practical applications, the self-learning task configuration information configured by the user for the self-learning task includes, but is not limited to: basic information, sample construction configuration information, fine-tuning training configuration information required for model training, and model evaluation configuration information, etc.

[0029] In this embodiment, basic information refers to information that is set on demand related to model updates, such as including but not limited to: model name, model function description, task triggering mechanism, and model release mechanism.

[0030] Task triggering mechanisms include, but are not limited to: triggering condition information, methods for obtaining tracking data, and dataset requirements.

[0031] Triggering condition information refers to the conditions that trigger the self-learning task, i.e., the conditions that trigger the model update event. This includes, but is not limited to, triggering cycle information and specified triggering event information. Triggering cycle information indicates the time interval for triggering the model update event, for example, triggering a model update event every 7, 15, 30, or 90 days. Specified triggering event information refers to information about pre-specified related events that trigger the model update event, including but not limited to, manually triggered events, specified abnormal events, or critical events.

[0032] The methods for obtaining event tracking feedback data include, but are not limited to: the storage path of the event tracking feedback data (also known as the retrieval address) and the time range corresponding to the event tracking feedback data. The time range corresponding to the event tracking feedback data indicates the time range within which event tracking feedback data is retrieved.

[0033] Dataset requirement information indicates the required sample dataset for model updates, including but not limited to: the percentage of training sample datasets needed for each model training session, the maximum amount of training sample data needed for each model training session, the maximum amount of historical test sample datasets needed for each model training session, and the historical time range corresponding to the historical test sample datasets. Historical test sample datasets refer to test sample datasets used for model testing within a historical time range. For example, historical test sample datasets from 30, 60, 90, or 180 days ago. In this embodiment, the amount of historical test sample datasets needed for each model training session and the corresponding historical time range can be considered as a method of obtaining historical test sample datasets. Of course, other methods can also be used to obtain historical test sample datasets, such as the storage path of the historical test sample datasets.

[0034] Model release mechanisms include, but are not limited to: manual release mechanisms that instruct algorithm models to be released manually and automatic release mechanisms that instruct algorithm models to be released automatically.

[0035] In this embodiment, the sample construction configuration information refers to the information configured for sample construction, such as, but not limited to, the screening strategy adopted by the active learning method.

[0036] In this embodiment, the fine-tuning training configuration information required for model training refers to the information configured for fine-tuning the algorithm model, including but not limited to: the model structure required for model update, the loss function required for model update, and the model parameters required for model update. The model structure required for model update, the loss function required for model update, and the model parameters required for model update can be understood as the training objects configured in the fine-tuning training configuration information. During model training, fine-tuning training is performed on the training objects configured in the fine-tuning training configuration information. Preferably, fine-tuning training based on incremental learning is performed on the training objects configured in the fine-tuning training configuration information. Incremental learning methods include, but are not limited to: regularization-based incremental learning, playback-based incremental learning, and parameter isolation-based incremental learning.

[0037] In this embodiment, model evaluation configuration information refers to the information configured for model evaluation, including but not limited to: performance test metrics and performance evaluation algorithms required for model evaluation.

[0038] Performance testing metrics include, but are not limited to: accuracy, precision, recall, F1 score (harmonic mean), coverage, and top-k accuracy. Top-k accuracy represents the matching of the k highest-scoring labels in a multi-class classification scenario with the true values. If even one label matches the true value, the result is true; otherwise, the result is false. k is a positive integer.

[0039] Performance evaluation algorithms refer to the execution algorithms used in comparative evaluation (which can also be understood as the technical logic that implements comparative evaluation functions), and can be flexibly defined as needed. In practical applications, one or more performance test metrics can be selected to participate in comparative evaluation. For example, the performance test metrics participating in the comparative evaluation are recall rate and F1 score, and the relationship between the performance test metrics is "AND". The performance evaluation algorithm is defined as follows: "If the recall rate of the new algorithm model (i.e., the updated algorithm model) is higher than the recall rate of the old algorithm model (i.e., the algorithm model before the update), and the F1 score of the new algorithm model is higher than the F1 score of the old algorithm model, the new algorithm model meets the conditions for online release; if the recall rate of the new algorithm model is lower than the recall rate of the old algorithm model, or if the F1 score of the new algorithm model is lower than the F1 score of the old algorithm model, the new algorithm model does not meet the conditions for online release."

[0040] For example, the performance test metrics used in the comparative evaluation are recall and F1 score. The relationship between the performance test metrics is "OR". The performance evaluation algorithm is: "If the recall of the new algorithm model is higher than that of the old algorithm model, or the F1 score of the new algorithm model is higher than that of the old algorithm model, the new algorithm model meets the conditions for online release; if the recall of the new algorithm model is lower than that of the old algorithm model, or the F1 score of the new algorithm model is lower than that of the old algorithm model, the new algorithm model does not meet the conditions for online release." Another example is: if the performance test metrics used in the comparative evaluation are accuracy, precision, recall, F1 score, coverage, and top-k accuracy (6 performance test metrics), the performance evaluation algorithm is: "If the new algorithm model (i.e., the updated algorithm model) has 4 performance test metrics higher than the old algorithm model (i.e., the algorithm model before the update), the new algorithm model meets the conditions for online release; otherwise, the new algorithm model does not meet the conditions for online release."

[0041] In practical applications, to improve the intelligence of self-learning task configuration, a self-learning task configuration interface is provided to users. Users can configure self-learning tasks as needed by interacting with this interface. Further optionally, to improve user-friendliness, see [link to relevant documentation]. Figure 2The self-learning task configuration interface may include one or more of the following: basic information configuration items, sample construction configuration items, model training configuration items, and model evaluation configuration items. Therefore, to meet the user's self-learning task configuration needs, the self-learning task configuration interface can be displayed in response to the user's input task configuration trigger operation. The self-learning task configuration interface includes at least one of the following: basic information configuration items, sample construction configuration items, model training configuration items, and model evaluation configuration items.

[0042] In this embodiment, the user operates on the basic information configuration items to configure the basic information required for model updates. See also Figure 2 Based on this, in response to configuration operations triggered by basic information configuration items, the basic information required for model updates is configured. Basic information may include, but is not limited to, a first acquisition method for obtaining tracking feedback data and / or a second acquisition method for obtaining historical test sample datasets.

[0043] In this embodiment, the user operates on the sample construction configuration item to configure the sample construction configuration information. Based on this, in response to the configuration operation triggered by the sample construction configuration item, the sample construction configuration information is configured.

[0044] Further, optionally, to meet the diverse sample construction needs of algorithm models across different paradigms, various filtering strategies can be provided for users to choose from. Based on this, see [link to relevant documentation]. Figure 2 In the process of configuring sample construction configuration information, if it is necessary to configure a target filtering strategy for active learning, then in response to the configuration operation triggered by the sample construction configuration item, one possible implementation of configuring sample construction configuration information is as follows: in response to the configuration operation triggered by the sample construction configuration item, display multiple filtering strategies; in response to the filtering strategy selection operation, configure the filtering strategy selected from the multiple filtering strategies as the target filtering strategy.

[0045] These filtering strategies include, but are not limited to: filtering strategies based on uncertainty sampling queries, filtering strategies based on query by committee queries, filtering strategies based on expected model change queries, filtering strategies based on expected error reduction queries, filtering strategies based on variance reduction queries, and filtering strategies based on density-weighted methods queries. Of course, filtering strategies can be flexibly defined as needed.

[0046] In this embodiment, the user operates on the model training configuration items to configure fine-tuning training configuration information. Based on this, in response to the configuration operation triggered by the model training configuration items, the fine-tuning training configuration information is configured. Further optional, see [link to relevant documentation]. Figure 2 To improve user-friendliness, model training configuration items may include at least one of the following: model structure sub-configuration item, loss function sub-configuration item, model parameter sub-configuration item, training method configuration item, and learning mode sub-configuration item.

[0047] In the process of configuring fine-tuning training configuration information, if it is necessary to configure the target model structure required for model update training, then in response to the configuration operation triggered by the model training configuration item, one possible implementation of configuring sample construction configuration information is: in response to the configuration operation triggered by the model structure sub-configuration item, configure the target model structure required for model update training.

[0048] In the process of configuring fine-tuning training configuration information, if it is necessary to configure the target loss function required for model update training, then in response to the configuration operation triggered by the model training configuration item, one optional implementation of configuring fine-tuning training configuration information is: in response to the configuration operation triggered by the loss function sub-configuration item, configure the target loss function required for model update training.

[0049] In the process of configuring fine-tuning training configuration information, if it is necessary to configure the target model parameters that need to be adjusted for model updates, then in response to the configuration operation triggered by the model training configuration item, one possible implementation of configuring fine-tuning training configuration information is: in response to the configuration operation triggered by the model parameter sub-configuration item, configure the target model parameters that need to be adjusted for model updates.

[0050] During the configuration of fine-tuning training configuration information, if it is necessary to configure the training method for model updates, one optional implementation for configuring fine-tuning training configuration information in response to configuration operations triggered by the model training configuration item is: in response to configuration operations triggered by the training method configuration item, configure the training method for model updates. Training methods include, but are not limited to, full-parameter training and partial-parameter training. Full-parameter training refers to updating all model parameters of the algorithm model, while partial-parameter training refers to updating only a portion of the model parameters out of the total model parameters.

[0051] During the configuration of fine-tuning training configuration information, if it is necessary to configure the learning mode of incremental learning, in response to the configuration operation triggered by the model training configuration item, one possible implementation of configuring fine-tuning training configuration information is as follows: in response to the configuration operation triggered by the learning mode sub-configuration item, display multiple learning modes; in response to the learning mode selection operation, configure the learning mode selected from the multiple learning modes as the target learning mode.

[0052] Among these, several learning modes include, but are not limited to: regularization-based incremental learning, playback-based incremental learning, and parameter isolation-based incremental learning.

[0053] In practical applications, the learning mode can be configured along with the training model structure, loss function, model parameters, or training method required for model updates. Alternatively, the learning mode can be configured independently, meaning it doesn't need to be configured along with the training model structure, loss function, model parameters, or training method for model updates; there are no restrictions on this.

[0054] In this embodiment, the user operates on the model evaluation configuration item to configure model evaluation configuration information. Based on this, in response to the configuration operation triggered by the model evaluation configuration item, the model evaluation configuration information required for model evaluation is configured. The model evaluation configuration information may include target performance test metrics and / or target performance evaluation algorithms, but is not limited to these.

[0055] Further optional, to improve user-friendliness, model evaluation configuration items may include, but are not limited to, performance test metric sub-configuration items and performance evaluation algorithm sub-configuration items.

[0056] In the process of configuring model evaluation configuration information, if it is necessary to configure target performance test metrics, then in response to the configuration operation triggered by the model evaluation configuration item, one optional implementation method for configuring the model evaluation configuration information required for model evaluation is: in response to the configuration operation triggered by the performance test metric sub-configuration item, configure the target performance test metrics.

[0057] In the process of configuring model evaluation configuration information, if it is necessary to configure a target performance evaluation algorithm, in response to the configuration operation triggered by the model evaluation configuration item, one optional implementation of configuring the model evaluation configuration information required for model evaluation is as follows: in response to the configuration operation triggered by the performance evaluation algorithm sub-configuration item, display multiple existing performance evaluation algorithms; in response to the selection operation for the existing performance evaluation algorithm, configure the existing performance evaluation algorithm selected from the multiple existing performance evaluation algorithms as the target performance evaluation algorithm.

[0058] Optionally, to better meet user needs, custom target performance evaluation algorithms are also supported. Based on this, in response to multiple existing performance evaluation algorithms failing to meet user evaluation requirements, a custom performance evaluation algorithm configuration page can be displayed; in response to configuration operations entered through the custom performance evaluation algorithm configuration page, a custom performance evaluation algorithm can be created and configured as the target performance evaluation algorithm. Optionally, to enrich existing performance evaluation algorithms, custom performance evaluation algorithms can also be saved as new existing performance evaluation algorithms. In practical applications, custom performance evaluation algorithms can be flexibly set as needed, without any restrictions.

[0059] Figure 3 This is a flowchart illustrating a self-learning method for updating an algorithm model, provided in an embodiment of this application. The method can be executed by an electronic device, such as a cloud server, but is not limited thereto.

[0060] See Figure 3 The method may include the following steps: 301. In response to the model update event being triggered, obtain the self-learning task configuration information. The self-learning task configuration information includes basic information, sample construction configuration information, and fine-tuning training configuration information. The basic information includes the first acquisition method used to obtain the tracking feedback data.

[0061] 302. Obtain the tracking feedback data according to the first acquisition method.

[0062] 303. Based on the tracking data and sample construction configuration information, construct samples using an active learning approach to obtain the training sample dataset required for this model update.

[0063] 304. Based on the fine-tuning training configuration information and the current training sample dataset, the algorithm model before this update is fine-tuned using an incremental learning approach to obtain the updated algorithm model.

[0064] In this embodiment, the model update event can be triggered manually or automatically, without limitation. For example, if the self-learning task configuration information includes conditions for triggering the model update event, the system can automatically monitor whether these conditions are met and, if so, automatically trigger the model update event.

[0065] In this embodiment, the self-learning task configuration information is obtained whenever a model update event is triggered. The configuration process for the self-learning task configuration information can be found in the relevant descriptions of the foregoing embodiments.

[0066] First, after the model update event is triggered, the first acquisition method for obtaining the event tracking feedback data is obtained from the basic information included in the self-learning task configuration information, and the event tracking feedback data is obtained according to the first acquisition method. The first acquisition method includes, but is not limited to, the storage path of the event tracking feedback data and the time range corresponding to the event tracking feedback data, etc.

[0067] Next, sample construction configuration information is obtained from the self-learning task configuration information. Based on the event tracking feedback data and the sample construction configuration information, sample construction is performed using an active learning approach to obtain the training sample dataset required for this model update. Here, the training sample dataset refers to the training sample dataset prepared for this model update. Event tracking feedback data refers to relevant information providing feedback on the online performance of the algorithm model.

[0068] Optionally, to improve the flexibility of sample construction, the user can configure a filtering strategy for the active learning method in the sample construction configuration information as needed. For ease of understanding and distinction, the filtering strategy configured for the active learning method is referred to as the target filtering strategy. Based on this, when constructing samples for the active learning method, the event tracking feedback data is filtered according to the target filtering strategy to obtain the filtered event tracking feedback data; the samples are then constructed based on the filtered event tracking feedback data to obtain the training sample dataset for this session.

[0069] Specifically, when filtering event tracking data based on an active learning approach according to a target filtering strategy, event tracking data whose classification results do not meet the criteria for accurate classification or whose prediction results do not meet the criteria for determinism can be filtered out.

[0070] A classification result that does not meet the accuracy criteria means the accuracy of the classification result is low, while a classification result that meets the accuracy criteria means the accuracy of the classification result is high. In practical applications, the accuracy criteria can be flexibly set as needed. For example, a probability threshold can be set to determine whether a classification result is accurate. If the probability corresponding to the classification result is greater than or equal to the probability threshold, the accuracy criteria are met; if the probability corresponding to the classification result is less than the probability threshold, the accuracy criteria are not met. For example, when classifying animals in an image, if the classification result indicates a 59% probability that the animal is a cat, this means the probability of it being a cat is low, and the accuracy of the classification result is low, failing to meet the accuracy criteria. If the classification result indicates a 99% probability that the animal is a cat, this means the probability of it being a cat is high, and the accuracy of the classification result is high, meeting the accuracy criteria.

[0071] A prediction that meets the deterministic condition has high reliability; a prediction that does not meet the deterministic condition has low reliability. In practical applications, the deterministic condition can be flexibly set as needed. For example, a probability threshold can be set to define whether a prediction is reliable. If the probability of the prediction is greater than or equal to the probability threshold, the deterministic condition is met; if the probability is less than the probability threshold, the deterministic condition is not met. For example, a prediction with a 59% probability of rain has high uncertainty and low reliability, thus not meeting the deterministic condition; a prediction with a 99% probability of rain has low uncertainty and high reliability, thus meeting the deterministic condition.

[0072] In this embodiment, when constructing the training sample dataset based on the filtered event tracking feedback data, the dataset requirement information included in the basic information of the self-learning task configuration information can be used as a reference. The selected event tracking feedback data can then be used to construct the training sample dataset that meets the dataset requirement information. For example, if the training sample data of the algorithm model includes model input data, model output results, and annotation results from the event tracking feedback data, and the dataset requirement information indicates that the maximum amount of training sample data required for each model training session is 10,000 records, then a maximum of 10,000 event tracking feedback data records can be selected from the filtered data. From each record, model input data, model output results, and annotation results are extracted, and these extracted data are used as one training sample data record.

[0073] In practical applications, considering that the feedback data from tracking points often contains noise or changes in data distribution, adopting an active learning approach in the sample construction stage can better select representative high-quality samples, effectively reduce the amount of sample data, and enhance the feasibility of the algorithm model in low-resource environments.

[0074] In this embodiment, during the fine-tuning training phase, fine-tuning training configuration information is obtained from the self-learning task configuration information. Based on the fine-tuning training configuration information and the current training sample dataset, the algorithm model before this update is fine-tuned using an incremental learning approach to obtain the updated algorithm model.

[0075] Specifically, the fine-tuning training configuration information can indicate the training object targeted by the fine-tuning training, and it can also indicate the learning mode of the incremental learning method. For ease of understanding and distinction, the learning mode of the incremental learning method involved in this model update is called the target learning mode.

[0076] In this embodiment, the training objects configured by the fine-tuning training configuration information are fine-tuned in an incremental learning manner according to the target learning mode. This avoids catastrophic forgetting of the model, ensures that the algorithm model can continuously learn knowledge from new data, and maintain its performance on historical data without decline. This enhances the algorithm model's ability to learn from changes in the customer's application data and its generalization performance.

[0077] In some optional embodiments, users are supported in modifying the model structure of the algorithm model and fine-tuning the modified model structure during training. When modifying the model structure, users can modify existing model structures or add various model structures such as convolutional layers, pooling layers, or fully connected layers; there are no restrictions on this. The model structure required for this model update is referred to as the target model structure. If the fine-tuning training configuration information indicates that the model update requires the target model structure, then the algorithm model before the update is fine-tuned using incremental learning based on the current training sample dataset, with the updated model parameters as the target. This allows training on a small number of model structures within the entire algorithm model, improving training efficiency and enhancing the feasibility of deploying the algorithm model in low-resource environments.

[0078] In some optional embodiments, users are supported in modifying the loss function of the algorithm model and fine-tuning the modified loss function. When modifying the loss function, users can modify existing loss functions or add new ones; there are no restrictions. The loss function required for this model update is referred to as the target loss function. If the fine-tuning training configuration information indicates that the model update requires the target loss function, then the model parameters are adjusted with the target loss function as the objective. Based on the current training sample dataset, the algorithm model before the update is fine-tuned using an incremental learning approach to obtain the updated algorithm model. This allows training on a small number of loss functions of the entire algorithm model, improving training efficiency and enhancing the feasibility of deploying the algorithm model in low-resource environments.

[0079] In some optional embodiments, users can adjust the model parameters of the algorithm model and fine-tune the loss function to be adjusted. For example, users can select which model parameters need to be adjusted as needed, such as selecting all or some model parameters to adjust. Here, the model parameters to be adjusted for the model update are referred to as the target model parameters. If the fine-tuning training configuration information indicates that the model update requires adjustment of the target model parameters, then with the adjustment of the target model parameters as the target, the algorithm model before the update is fine-tuned using an incremental learning approach based on the current training sample dataset to obtain the updated algorithm model. In this way, the entire algorithm model can be trained with the selected model parameters as needed, improving training efficiency and enhancing the feasibility of the algorithm model in low-resource environments.

[0080] In some optional embodiments, users can configure the learning mode of incremental learning as needed. If the fine-tuning training configuration information indicates the target learning mode of incremental learning, then the algorithm model before the current update is fine-tuned based on incremental learning according to the current training sample dataset according to the target learning mode to obtain the updated algorithm model.

[0081] For example, with the goal of updating the model parameters of the target model structure, the algorithm model before the update is fine-tuned using an incremental learning approach based on the current training sample dataset, according to the target learning mode, to obtain the updated algorithm model. As another example, with the goal of adjusting the model parameters using the target loss function, the algorithm model before the update is fine-tuned using an incremental learning approach based on the current training sample dataset, according to the target learning mode, to obtain the updated algorithm model. And again, with the goal of adjusting the target model parameters, the algorithm model before the update is fine-tuned using an incremental learning approach based on the current training sample dataset, according to the target learning mode, to obtain the updated algorithm model.

[0082] The technical solution provided in this application forms a self-learning closed-loop chain encompassing data collection based on tracking points, sample construction based on active learning, and fine-tuning training based on incremental learning. This supports automated iteration of algorithm model updates for various paradigms, reducing the manual maintenance costs of algorithm model updates and improving update efficiency and intelligence. Furthermore, sample construction based on active learning can better select representative, high-quality samples, effectively reducing the amount of sample data, decreasing computational resource requirements, and enhancing the feasibility of deploying the algorithm model in low-resource environments. Fine-tuning training based on incremental learning avoids catastrophic forgetting, ensuring the algorithm model can continuously learn from new data while maintaining its performance on historical data, thus enhancing the algorithm model's ability to learn from changes in customer application data and its generalization performance.

[0083] Figure 4 This is a flowchart illustrating another self-learning method for updating an algorithm model provided in an embodiment of this application. The method can be executed by an electronic device, such as a cloud server, but is not limited thereto.

[0084] See Figure 4 The method may include the following steps: 401. In response to the model update event being triggered, obtain the self-learning task configuration information. The self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information, and model evaluation configuration information. The basic information includes the first acquisition method for obtaining the tracking feedback data and the second acquisition method for obtaining the historical test sample dataset. The model evaluation configuration information includes the target performance test metrics.

[0085] 402. Obtain the tracking feedback data according to the first acquisition method and obtain the historical test sample dataset according to the second acquisition method.

[0086] 403. Based on the tracking data and sample construction configuration information, construct samples using an active learning approach to obtain the training sample dataset and test sample dataset required for this model update.

[0087] 404. Based on the fine-tuning training configuration information and the current training sample dataset, the algorithm model before this update is fine-tuned using an incremental learning approach to obtain the updated algorithm model.

[0088] It is worth noting that the implementation of step 404 can refer to the implementation of step 304 in the aforementioned embodiments, and there is no limitation thereto.

[0089] 405. Using the historical test sample dataset and the current test sample dataset, perform performance tests on the updated algorithm model against the target performance test metrics to obtain the performance test results of the updated algorithm model.

[0090] It is worth noting that the implementation methods of steps 403 and 404 can refer to the implementation methods of steps 303 and 304 in the foregoing embodiments, and there are no restrictions on them.

[0091] In this embodiment, the ability to perform performance testing on the updated algorithm model is also supported. To this end, users can configure a second method for obtaining historical test sample datasets when configuring the self-learning task configuration information. This allows the historical test sample dataset to be obtained during the model update phase using the second method. Furthermore, when constructing samples based on active learning, in addition to constructing the current training sample dataset required for this model update, a new test sample dataset is also constructed. This new test sample dataset is the newly added test sample dataset required for this model update.

[0092] In practical applications, the test sample dataset created each time the model is updated can be saved as a new historical test sample dataset. For example, if the dataset requirement information in the self-learning configuration specifies a historical time range of 30 days for the historical test sample dataset, then if a model update event is triggered on a given day, the historical test sample dataset from the previous 30 days will be retrieved.

[0093] If the dataset requirement information also includes the percentage of training sample datasets needed for each model training session, the maximum amount of training sample data needed for each model training session, and the maximum amount of historical test sample datasets needed for each model training session, then during sample construction, the percentage of the current training sample dataset in the total sample dataset must satisfy the percentage information in the dataset requirement information. The total sample dataset includes the current training sample dataset, the current test sample dataset, and the historical test sample datasets. Furthermore, from the filtered event tracking feedback data, at most the training sample dataset and the historical test sample dataset that satisfy the maximum amount of data required in the dataset requirement information should be selected.

[0094] In this embodiment, when configuring the self-learning task configuration information, the user can configure the model evaluation configuration information as needed. Here, the performance test metrics configured in the model evaluation configuration information for performance testing are referred to as the target performance test metrics.

[0095] In this embodiment, the updated algorithm model can be tested against the target performance metrics using historical test sample datasets to obtain the performance test results. Alternatively, the updated algorithm model can be tested against the target performance metrics using the current test sample dataset to obtain the performance test results. Further optionally, to better ensure that the algorithm model can continuously learn knowledge from new data while maintaining its performance on historical data, the updated algorithm model can be tested against the target performance metrics using both historical and current test sample datasets to obtain the performance test results. Performance test results include, but are not limited to, accuracy, precision, recall, F1 score, coverage, top-k accuracy, etc.

[0096] The performance test results of the updated algorithm model can help determine whether it can be released online. If the performance test results indicate that the updated algorithm model has passed the performance test, it can be released online; if the results indicate that it has failed the performance test, it cannot be released online, and the algorithm model in the online environment will remain the same as before the update. In practical applications, the criteria for passing the performance test can be set as needed. For example, if the performance test results of multiple performance metrics of the updated algorithm model are greater than preset metric thresholds, the updated algorithm model has passed the performance test; otherwise, it has failed. Metric thresholds include, but are not limited to: accuracy threshold, precision threshold, recall threshold, F1 score threshold, coverage threshold, and top-k accuracy threshold, etc. It is worth noting that more information on using test sample datasets to perform performance testing on algorithm models can be found in related technical documents, which will not be elaborated here.

[0097] The technical solution provided in this application forms a self-learning closed-loop chain encompassing data collection based on event tracking, sample construction based on active learning, fine-tuning training based on incremental learning, and model performance testing. This supports automated iteration and updates of algorithm models across various paradigms, reducing manual maintenance costs and improving update efficiency and intelligence. Furthermore, the active learning-based sample construction better selects representative, high-quality samples, effectively reducing the amount of sample data, decreasing computational resource requirements, and enhancing the feasibility of deploying the algorithm model in low-resource environments. Additionally, the incremental learning-based fine-tuning training avoids catastrophic forgetting, ensuring the algorithm model continuously learns from new data while maintaining its performance on historical data, enhancing its ability to learn from changes in customer application data and its generalization performance. Finally, the provision of model performance testing capabilities provides a solid foundation for deciding whether to release the updated algorithm model online, improving its intelligence.

[0098] Figure 5 This is a flowchart illustrating another self-learning method for updating an algorithm model provided in an embodiment of this application. The method can be executed by an electronic device, such as a cloud server, but is not limited thereto.

[0099] See Figure 5 The method may include the following steps: 501. In response to the model update event being triggered, obtain the self-learning task configuration information. The self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information, and model evaluation configuration information. The basic information includes the first acquisition method for obtaining the tracking feedback data and the second acquisition method for obtaining the historical test sample dataset. The model evaluation configuration information includes the target performance test index and the target performance evaluation algorithm.

[0100] 502. Obtain the tracking feedback data according to the first acquisition method and obtain the historical test sample dataset according to the second acquisition method.

[0101] 503. Based on the tracking data and sample construction configuration information, construct samples using an active learning approach to obtain the training sample dataset and test sample dataset required for this model update.

[0102] 504. Based on the fine-tuning training configuration information and the current training sample dataset, the algorithm model before this update is fine-tuned using an incremental learning approach to obtain the updated algorithm model.

[0103] 505. Using the historical test sample dataset and the current test sample dataset, perform performance tests on the updated algorithm model against the target performance test metrics to obtain the performance test results of the updated algorithm model.

[0104] 506. Using the historical test sample dataset and the current test sample dataset, perform performance tests on the algorithm model before this update against the target performance test metrics to obtain the performance test results of the algorithm model before this update.

[0105] 507. Using the target performance evaluation algorithm, compare and evaluate the performance test results of the algorithm model before the update and the algorithm model after the update to obtain the comparison evaluation results.

[0106] 508. Determine whether the updated algorithm model meets the conditions for going live based on the comparative evaluation results.

[0107] In this embodiment, the ability to compare and evaluate the algorithm model before and after the update is also supported. To this end, when configuring the model evaluation configuration information, the user can configure not only the target performance test indicators but also the performance evaluation algorithm for the model evaluation. Here, the performance evaluation algorithm configured for the model evaluation is referred to as the target performance evaluation algorithm. The implementation methods for steps 502, 503, 504, and 505 can be found in the foregoing embodiments and will not be repeated here.

[0108] In this embodiment, in addition to performing performance tests on the updated algorithm model, performance tests are also conducted on the algorithm model before the update. The target performance evaluation algorithm is used to compare the performance test results of the algorithm model before and after the update, resulting in a comparative evaluation result. Based on the comparative evaluation result, it is determined whether the updated algorithm model meets the conditions for online operation. For details regarding comparative evaluation, please refer to the previous section on comparative evaluation; it will not be repeated here.

[0109] If the evaluation results indicate that the performance of the updated online algorithm model is better than that of the original online algorithm model, then the updated online algorithm model can be deployed online. In this case, the updated online algorithm model is deployed to the online environment to complete the model update task in the online environment. If the evaluation results indicate that the performance of the updated online algorithm model is worse than that of the original online algorithm model, then the updated online algorithm model cannot be deployed online. In this case, the algorithm model in the online environment remains the original algorithm model.

[0110] Considering that customers may deploy multiple algorithm models in their service pipeline, they are more concerned with the overall processing performance of the entire pipeline than the performance metrics of individual algorithm models. Therefore, we can also provide a comprehensive evaluation capability for multiple algorithm models, which can be called end-to-end evaluation capability, to better meet customer needs and improve the user experience of algorithm models.

[0111] Based on the above, further optionally, if there are multiple algorithm models, these multiple algorithm models form an evaluation link; an optional implementation method for comparing and evaluating the performance test results of the algorithm models before and after the current update using the target performance evaluation algorithm is as follows: based on the correlation between the algorithm models, the target performance evaluation algorithm determines the comprehensive performance test result of the evaluation link before the update based on the performance test results of the multiple algorithm models before the update; based on the correlation between the algorithm models, the target performance evaluation algorithm determines the comprehensive performance test result of the evaluation link after the update based on the performance test results of the multiple algorithm models after the update; and the comprehensive performance test results before and after the update are compared and evaluated.

[0112] In this embodiment, the link composed of multiple algorithm models is called the evaluation link. The relationships between algorithm models on the evaluation link include, but are not limited to, serial, parallel, and competitive relationships. If the algorithm models are serially related, the performance test results of both serially related algorithm models are included in the overall performance measurement result. If the algorithm models are parallelly related, either of the performance test results of the two parallel algorithm models can be selected to participate in the overall performance measurement result. If the algorithm models are competitive, only the performance test result with a competitive advantage from the two competitive algorithm models can be included in the overall performance measurement result. In practical applications, the target performance evaluation algorithm can define the processing algorithm for the performance test results of algorithm models with different relationships according to the actual application requirements, and there are no restrictions on this. Figure 6 For example, the evaluation chain includes model A and model B. If the relationship between model A and model B is serial, the performance test results of model A and model B can be multiplied or added together to obtain the comprehensive performance test result of the evaluation chain. If the relationship between model A and model B is parallel, the performance test result of either model A or model B can be selected as the comprehensive performance test result of the evaluation chain. If the relationship between model A and model B is competitive, the performance test result of model B, which has a competitive advantage, can be selected as the comprehensive performance test result of the evaluation chain.

[0113] The technical solution provided in this application forms a self-learning closed-loop chain encompassing data collection based on event tracking, sample construction based on active learning, fine-tuning training based on incremental learning, and comparative evaluation of the algorithm models before and after updates. This supports automated iteration of algorithm model updates for various paradigms, reducing the manual maintenance costs of algorithm model updates and improving update efficiency and intelligence. Furthermore, sample construction based on active learning can better select representative, high-quality samples, effectively reducing the amount of sample data, decreasing computational resource requirements, and enhancing the feasibility of deploying the algorithm model in low-resource environments. Furthermore, fine-tuning training based on incremental learning avoids catastrophic forgetting, ensuring the algorithm model can continuously learn from new data while maintaining its performance on historical data, enhancing its learning ability and generalization performance in response to changes in customer application data. Finally, providing comparative evaluation capabilities between the algorithm models before and after updates provides a solid foundation for deciding whether to release the updated algorithm model online, improving the intelligence of the algorithm model.

[0114] Figure 7 This is a schematic diagram illustrating the structure of a self-learning system for updating an algorithm model, provided in an embodiment of this application. This system can be deployed on a cloud server. (See attached diagram.) Figure 7 The system may include a control module 20, and a human-computer interaction module 10, a sample construction module 30, a model training module 40 and a model evaluation module 50, which are respectively connected to the control module 20. The human-computer interaction module 10 is used to respond to the user's input task configuration trigger operation and display the self-learning task configuration interface; in response to the configuration operation triggered by the user through the self-learning task configuration interface, it configures the self-learning task configuration information and sends the self-learning task configuration information to the control module 20. The self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information and model evaluation configuration information. The basic information includes a first acquisition method for obtaining the tracking feedback data and a second acquisition method for obtaining the historical test sample dataset. The model evaluation configuration information includes the target performance test index and the target performance evaluation algorithm. The sample construction module 30 is used to acquire the tracking feedback data according to the first acquisition method under the control of the control module 20, and to construct samples based on the active learning method according to the tracking feedback data and the sample construction configuration information to obtain the training sample dataset and the test sample dataset required for this model update. The model training module 40 is used, under the control of the control module 20, to fine-tune the algorithm model before this update based on incremental learning according to the fine-tuning training configuration information and the current training sample dataset, so as to obtain the updated algorithm model. The model evaluation module 50, under the control of the control module 20, acquires a historical test sample dataset according to the second acquisition method; uses the historical test sample dataset and the current test sample dataset to perform performance tests on the algorithm model before the update and the algorithm model after the update, respectively, targeting the target performance test indicators; uses the target performance evaluation algorithm to compare and evaluate the performance test results of the algorithm model before the update and the algorithm model after the update, respectively, to obtain the comparison evaluation results; and determines whether the updated algorithm model meets the conditions for online operation based on the comparison evaluation results.

[0115] The implementation methods of each module in the system embodiment have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0116] Specifically, see Figure 7 As shown in ①, users can initiate task configuration trigger operations as needed. The human-computer interaction module 10 responds to the user's input of the task configuration trigger operation and displays the self-learning task configuration interface; in response to the configuration operation triggered by the user through the self-learning task configuration interface, it configures the self-learning task configuration information, see [link to module 10]. Figure 7 As shown in ②, the human-computer interaction module 10 sends self-learning task configuration information to the control module 20. (See also...) Figure 7 As shown in ③, the self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information, and model evaluation configuration information. Control module 20 controls sample construction module 30, model training module 40, and model evaluation module 50 to perform corresponding operations based on the self-learning task configuration information. See also... Figure 7 As shown in ④, the control module 20 triggers the sample construction module 30 to construct samples based on active learning according to the self-learning task configuration information. See also Figure 7 As shown in ⑤, the control module 20 triggers the model training module 40 to perform fine-tuning training based on incremental learning according to the self-learning task configuration information. See also Figure 7 As shown in step ⑥, the control module 20 triggers the model evaluation module 50 to perform a comparative evaluation according to the self-learning task configuration information. This completes the self-learning closed-loop chain.

[0117] The self-learning system for updating algorithm models provides a standardized framework for model updates. Users can easily and quickly connect their algorithm models to the self-learning system for updating algorithm models, enabling automatic iterative updates and saving a significant amount of human resources.

[0118] The self-learning system for updating algorithm models provided in this application forms a closed-loop self-learning chain, encompassing data collection based on tracking points, sample construction based on active learning, fine-tuning training based on incremental learning, and comparative evaluation between the updated and unupdated algorithm models. This system supports automated iteration of algorithm model updates for various paradigms, reducing manual maintenance costs and improving update efficiency and intelligence. Furthermore, the active learning-based sample construction better selects representative, high-quality samples, effectively reducing the amount of sample data, decreasing computational resource requirements, and enhancing the feasibility of deploying the algorithm model in low-resource environments. Additionally, the incremental learning-based fine-tuning training avoids catastrophic forgetting, ensuring the algorithm model continuously learns from new data while maintaining its performance on historical data, enhancing its learning ability and generalization performance in response to changes in customer application data. Finally, the system provides comparative evaluation capabilities between the updated and unupdated algorithm models, providing a solid foundation for deciding whether to release the updated model and improving its intelligence.

[0119] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 301 to 304 can be device A; or the execution subject of steps 301 and 302 can be device A, and the execution subject of steps 303 and 304 can be device B; and so on.

[0120] Furthermore, in some of the processes described in the above embodiments and accompanying drawings, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 301, 302, etc., are merely used to distinguish different operations and do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0121] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0122] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device includes: a memory 81 and a processor 82; Memory 81 is used to store computer programs and can be configured to store various other data to support operation on the computing platform. Examples of this data include instructions for any application or method operating on the computing platform, contact data, phone book data, messages, pictures, videos, etc.

[0123] The memory 81 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random-access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0124] Processor 82, coupled to memory 81, is used to execute computer programs in memory 81 for performing steps in a self-learning method to update the algorithm model.

[0125] Furthermore, such as Figure 8 As shown, the electronic device also includes other components such as a communication component 83, a display 84, a power supply component 85, and an audio component 86. Figure 8 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 8 The components shown. Additionally... Figure 8 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the electronic device. The electronic device in this embodiment can be a desktop computer, laptop computer, smartphone, or IoT (Internet of Things) device, or a server-side device such as a conventional server, cloud server, or server array. If the electronic device in this embodiment is a desktop computer, laptop computer, or smartphone, it may include... Figure 8 The components within the dashed box; if the electronic device in this embodiment is implemented as a conventional server, cloud server, or server array, it may be omitted. Figure 8The component within the dashed box.

[0126] For a detailed description of the implementation process of each action by the processor, please refer to the relevant descriptions in the foregoing method embodiments or device embodiments, which will not be repeated here.

[0127] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed, can implement the steps that can be performed by an electronic device in the above method embodiments.

[0128] Accordingly, this application also provides a computer program product, including a computer program / instructions, which, when executed by a processor, enables the processor to perform the steps that can be executed by an electronic device in the above method embodiments.

[0129] The aforementioned communication components are configured to facilitate wired or wireless communication between the device containing the communication components and other devices. The device containing the communication components can access wireless networks based on communication standards, such as WiFi (Wireless Fidelity), 2G (2nd Generation), 3G (3rd Generation), 4G (4th Generation) / LTE (long Term Evolution), 5G (5th Generation), or combinations thereof. In one exemplary embodiment, the communication components receive broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, the communication components also include a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module may be based on Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra Wide Band (UWB), Bluetooth, and other technologies.

[0130] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0131] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0132] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0133] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0137] In a typical configuration, a computing device includes one or more processors (Central Processing Unit, CPU), input / output interfaces, network interfaces, and memory.

[0138] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0139] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change RAM (PRAM), static random-access memory (SRAM), dynamic random-access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined in this article, computer-readable media do not include transient media, such as modulated data signals and carrier waves.

[0140] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0141] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A self-learning method for updating an algorithm model, characterized in that, The method, executed by a cloud provider offering cloud computing resources, includes: In response to a user-inputted task configuration trigger operation, a self-learning task configuration interface is displayed, which includes basic information configuration items, sample construction configuration items, model training configuration items, and model evaluation configuration items. In response to the configuration operation triggered by the user through the self-learning task configuration interface, configure the self-learning task configuration information; In response to a model update event being triggered, the self-learning task configuration information is obtained. The self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information, and model evaluation configuration information. The basic information includes a first acquisition method for obtaining tracking feedback data and a second acquisition method for obtaining historical test sample datasets. The model evaluation configuration information includes target performance test metrics and target performance evaluation algorithms. The tracking feedback data is obtained according to the first acquisition method. The tracking feedback data refers to relevant information that provides feedback on the online running effect of the algorithm model. The tracking feedback data includes model input data, model output results, and annotation results. Based on the data collected from the tracking points and the sample construction configuration information, a sample construction based on active learning is performed to obtain the training sample dataset required for this model update. Based on the fine-tuning training configuration information and the current training sample dataset, the algorithm model before this update is fine-tuned using an incremental learning approach to obtain the updated algorithm model. The historical test sample dataset is obtained according to the second acquisition method; based on the tracking feedback data and the sample construction configuration information, a sample is constructed using an active learning approach to obtain the current test sample dataset required for this model update; the updated algorithm model is then tested against the target performance test metrics using the historical test sample dataset and the current test sample dataset to obtain the performance test results of the updated algorithm model. The target performance evaluation algorithm is used to compare and evaluate the performance test results of the algorithm model before the update and the algorithm model after the update, and the comparison evaluation results are obtained. Based on the comparative evaluation results, determine whether the updated algorithm model meets the conditions for online operation; if it meets the conditions for online operation, then release the updated algorithm model to the online environment to complete the model update task of the algorithm model in the online environment.

2. The method according to claim 1, characterized in that, The sample construction configuration information includes the target selection strategy of the active learning method; Based on the data tracking feedback and the sample construction configuration information, a sample construction based on active learning is performed to obtain the training sample dataset required for this model update, including: The tracking feedback data is filtered based on an active learning method according to the target filtering strategy to obtain the filtered tracking feedback data. The selected tracking data is used to construct a sample dataset for this training exercise.

3. The method according to claim 1 or 2, characterized in that, Based on the fine-tuning training configuration information and the current training sample dataset, the algorithm model before this update is fine-tuned using an incremental learning approach to obtain the updated algorithm model, including: If the fine-tuning training configuration information indicates that the model needs to update the target model structure to be trained, then with the model parameters of the target model structure being updated as the target, the algorithm model before the current update is fine-tuned based on incremental learning according to the current training sample dataset to obtain the algorithm model after the current update. And / or, If the fine-tuning training configuration information indicates that the target loss function required for model update is to be trained, then the algorithm model before this update is fine-tuned based on incremental learning according to the current training sample dataset, with the target loss function as the objective, to obtain the updated algorithm model. And / or, If the fine-tuning training configuration information indicates the target model parameters that need to be adjusted for the model update, then with the adjustment of the target model parameters as the goal, the algorithm model before the update is fine-tuned based on incremental learning according to the current training sample dataset to obtain the updated algorithm model.

4. The method according to claim 1, characterized in that, If there are multiple algorithm models, the multiple algorithm models form an evaluation link; The target performance evaluation algorithm is used to compare and evaluate the performance of the algorithm model before and after the update, respectively, to obtain the comparison and evaluation results, including: Based on the correlation between algorithm models, the comprehensive performance test result of the evaluation link before the update is determined using the target performance evaluation algorithm based on the performance test results of multiple algorithm models before this update. Based on the correlation between algorithm models, the comprehensive performance test result of the updated evaluation link is determined using the target performance evaluation algorithm based on the performance test results of multiple updated algorithm models. The comprehensive performance test results of the evaluation link before and after the update are compared and evaluated.

5. The method according to any one of claims 1 to 2, characterized in that, In response to the configuration operation triggered by the user through the self-learning task configuration interface, the configuration information for the self-learning task includes: In response to a user-inputted task configuration trigger operation, a self-learning task configuration interface is displayed, which includes basic information configuration items, sample construction configuration items, model training configuration items, and model evaluation configuration items. In response to a configuration operation triggered by a basic information configuration item, the basic information required for the configuration model update is configured, wherein the basic information includes the first acquisition method and the second acquisition method; In response to a configuration operation triggered by a sample build configuration item, configure the sample build configuration information; In response to a configuration operation triggered by the model training configuration item, configure the fine-tuning training configuration information; In response to a configuration operation triggered by the model evaluation configuration item, the model evaluation configuration information required for model evaluation is configured, including target performance test metrics and target performance evaluation algorithms.

6. The method according to claim 5, characterized in that, The sample construction configuration information includes at least: a target selection strategy configured for active learning; In response to a configuration operation triggered by a sample build configuration item, configure the sample build configuration information, including: In response to configuration actions triggered by constructing configuration items from samples, multiple filtering strategies are displayed; In response to the filtering strategy selection operation, the filtering strategy selected from the plurality of filtering strategies is configured as the target filtering strategy.

7. The method according to claim 5, characterized in that, The model training configuration items include: model structure sub-configuration items, loss function sub-configuration items, model parameter sub-configuration items, and learning mode sub-configuration items; In response to a configuration operation triggered by the model training configuration item, configure the fine-tuning training configuration information, including: In response to configuration operations triggered by the model structure sub-configuration item, configure the target model structure required for training to be updated. In response to configuration operations triggered by the loss function sub-configuration item, configure the target loss function required for model updates. In response to configuration operations triggered by model parameter sub-configuration items, configure the target model parameters that need to be adjusted for model updates; In response to configuration actions triggered by the learning mode sub-configuration item, display multiple learning modes; In response to the learning mode selection operation, the learning mode selected from the plurality of learning modes is configured as the target learning mode.

8. The method according to claim 5, characterized in that, The model evaluation configuration items include performance test index sub-configuration items and performance evaluation algorithm sub-configuration items; In response to a configuration operation triggered by the model evaluation configuration item, configure the target performance test metrics and target performance evaluation algorithm required for model evaluation, including: In response to a configuration operation triggered by the performance test metric sub-configuration item, configure the target performance test metric; In response to a configuration operation triggered by the performance evaluation algorithm sub-configuration item, multiple existing performance evaluation algorithms are displayed; In response to the selection operation for an existing performance evaluation algorithm, an existing performance evaluation algorithm selected from multiple existing performance evaluation algorithms is configured as the target performance evaluation algorithm.

9. The method according to claim 8, characterized in that, Also includes: In response to the fact that several existing performance evaluation algorithms do not meet the user's evaluation needs, the custom performance evaluation algorithm configuration page is displayed; In response to the configuration operation input through the custom performance evaluation algorithm configuration page, a custom performance evaluation algorithm is created and configured as the target performance evaluation algorithm; as well as The custom performance evaluation algorithm is saved as a new existing performance evaluation algorithm.

10. A self-learning system for updating an algorithm model, characterized in that, The system is deployed on a cloud server and includes: a control module, and a human-computer interaction module, a sample construction module, a model training module, and a model evaluation module, which are respectively connected to the control module; The human-computer interaction module is used to display a self-learning task configuration interface in response to a user-input task configuration trigger operation; in response to a configuration operation triggered by the user through the self-learning task configuration interface, it configures self-learning task configuration information and sends the self-learning task configuration information to the control module. The self-learning task configuration information includes basic information, sample construction configuration information, fine-tuning training configuration information, and model evaluation configuration information. The basic information includes a first acquisition method for obtaining tracking feedback data and a second acquisition method for obtaining historical test sample datasets. The model evaluation configuration information includes target performance test metrics and target performance evaluation algorithms. The sample construction module is used, under the control of the control module, to acquire the tracking feedback data according to the first acquisition method, and to construct samples based on active learning according to the tracking feedback data and the sample construction configuration information to obtain the current training sample dataset and the current test sample dataset required for this model update; the tracking feedback data refers to relevant information that provides feedback on the online running effect of the algorithm model; the tracking feedback data includes model input data, model output results, and annotation results; The model training module is used to perform fine-tuning training on the algorithm model before this update based on incremental learning, under the control of the control module, according to the fine-tuning training configuration information and the current training sample dataset, to obtain the updated algorithm model. The model evaluation module is configured to, under the control of the control module, acquire the historical test sample dataset according to the second acquisition method; perform performance tests on the algorithm model before the current update and the algorithm model after the current update against the target performance test indicators using the historical test sample dataset and the current test sample dataset, respectively; compare and evaluate the performance test results of the algorithm model before the current update and the algorithm model after the current update using the target performance evaluation algorithm to obtain a comparison evaluation result; determine whether the algorithm model after the current update meets the conditions for online operation based on the comparison evaluation result; if the conditions for online operation are met, release the algorithm model after the current update to the online environment to complete the model update task of the algorithm model in the online environment.

11. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is coupled to the memory for executing the computer program to perform the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method according to any one of claims 1-9.

Images