A method for improving model performance by using incremental learning

Abstract

The application discloses a method for improving model performance by using incremental learning, comprising the following steps: constructing a historical core data set based on historical training data; performing desensitization processing on the historical core data set to obtain a historical playback data set; collecting false alarm data in the model running process, and labeling the false alarm data to obtain a new scene training data set; combining the new scene training data set and the historical playback data set to obtain an incremental training data set, and performing incremental training on a target model based on a preset training strategy; and outputting an optimized target model based on the model parameters after the incremental training is completed.

Description

Technical Field

[0001] This invention is applied to the field of artificial intelligence, specifically a method for improving model performance using the incremental learning approach. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning models have achieved remarkable results in fields such as image recognition and object detection. However, traditional deep learning often exhibits unstable performance when faced with constantly changing scenarios, especially when the actual application scenario differs significantly from the training scenario, resulting in a sharp drop in model performance and serious false positives and false negatives. In this case, the traditional method is to collect a large amount of new scenario data and combine it with existing historical training data to retrain the model. However, this not only consumes a lot of computing resources, but also, for industries that pay close attention to data privacy protection, the new scenario data cannot be retrieved, and the historical training data cannot be provided externally, thus hindering the implementation of the traditional training method based on large amounts of data.

[0003] The use of incremental learning to address traditional training methods has gradually gained attention. However, traditional incremental learning faces two major challenges: 1. Catastrophic forgetting, where old knowledge drops sharply after training on new data; 2. Privacy and storage, where storing all historical data carries high compliance risks and costs. This invention proposes a method and system for improving model performance based on incremental learning and a small amount of incremental data. By optimizing training strategies, it effectively reduces the model's forgetting rate, improves training efficiency, and enhances the overall performance of the model. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for improving model performance by using the incremental learning concept, which addresses the shortcomings of the existing technology.

[0005] To address the aforementioned technical problems, this invention provides a method for improving model performance using incremental learning, comprising the following steps:

[0006] Construct a historical core dataset based on historical training data;

[0007] The historical core dataset is anonymized to obtain the historical playback dataset;

[0008] Collect false alarm data during model operation and label the false alarm data to obtain a new scene training dataset;

[0009] The new scene training dataset is combined with the historical playback dataset to obtain an incremental training dataset, and the target model is incrementally trained based on a preset training strategy.

[0010] Based on the model parameters after the incremental training is completed, the optimized target model is output.

[0011] As one possible implementation, the steps for constructing a historical core dataset based on historical training data further include:

[0012] Obtain image data for each sample in the historical training dataset;

[0013] The image data is understood using a visual multimodal large model, and corresponding image text descriptions are generated based on preset prompts related to the target task.

[0014] The image text description is vectorized using a vision-language pre-trained model to obtain the feature vector of each image sample;

[0015] Clustering is performed based on the feature vectors to divide the historical training data into several clusters. The sample with the smallest distance to the feature vector of the cluster center is selected from each cluster as the representative sample of the cluster to form the historical core dataset.

[0016] As one possible implementation, the step of constructing a historical core dataset based on historical training data further includes:

[0017] While selecting representative samples based on feature vector clustering, samples are also selected from the historical training data and incorporated into the historical core dataset using at least one of the following methods:

[0018] A random sampling method is used to randomly select a preset number of samples from the historical training data;

[0019] The method for selecting difficult samples involves using existing models to predict historical training samples, sorting samples that are correctly predicted but have low confidence by category, and selecting a preset number of difficult samples from each category.

[0020] Based on the selection method of sample prediction uncertainty, different network types of models are used to predict the same sample for the same task. The difference in the prediction confidence of different models that predict the same category is used as the uncertainty index, and a preset number of samples are selected from the samples with larger differences.

[0021] As one possible implementation, the target tasks further include pedestrian detection tasks and / or vehicle detection tasks.

[0022] For the pedestrian detection task, the prompt words are used to guide the visual multimodal large model to output an image text description that includes the following:

[0023] The location, appearance, posture, behavior, and relationships between pedestrians or with the environment of each pedestrian;

[0024] For vehicle detection tasks, the cue words are used to guide the large visual multimodal model:

[0025] Determine whether a vehicle exists in the image. If a vehicle exists, output a text description for each vehicle, including at least one of the following: region location, vehicle type, color or paint scheme, orientation, degree of occlusion or truncation, abnormal state, environmental context, and potential detection challenges.

[0026] As one possible implementation, the preset training strategy further includes:

[0027] The parameters of the historical model, which are pre-trained based on the full historical training data, are used as the initialization parameters of the target model.

[0028] In the initial stage of training, the backbone network layer of the target model is frozen, and only the parameters of the neck network layer and the head network layer are updated. In the later stage of training, the backbone network layer is unfrozen to perform joint updates.

[0029] During each training iteration, a preset number of core samples are selected from the historical playback dataset and combined with samples from the new scene training dataset to form the current training batch.

[0030] As one possible implementation, the visual multimodal large model is further defined as at least one industry-specific visual multimodal large model, including any one or more of Tongyi Qianwen's Qwen2-VL or Qwen2.5-VL, OpenAI's GPT-4V, etc.

[0031] As one possible implementation, the vision-language pre-trained model is further selected from at least one of CLIP, BLIP-2, ViLT, and LLaVA.

[0032] As one possible implementation, the clustering is further performed using the K-means clustering algorithm, where the number of clusters is equal to the number of target historical core samples, and only one representative sample from each cluster is selected to be added to the historical core dataset.

[0033] As one possible implementation, the preset training strategy further includes:

[0034] During training, a learning rate that gradually decreases with the number of iterations is used to improve the model's convergence stability and reduce the forgetting of historical knowledge.

[0035] As a possible implementation, further, in each training iteration, in the step of selecting a preset number of core samples from the historical replay dataset and samples from the new scene training dataset to form the current training batch, the samples of the current training batch are composed of samples from the new scene training dataset and samples from the historical replay dataset in a preset ratio. The preset ratio is used to balance the adaptability of the new scene and the ability to retain historical knowledge during the training process.

[0036] The present invention adopts the above technical solution and has the following beneficial effects:

[0037] This invention is based on the idea of ​​incremental learning and proposes a method to improve model performance based on a small amount of incremental data. By extracting and optimizing the training strategy from historical core datasets and adding online false positive data, the optimized model can not only adapt to new scenarios, but also retain as much of the model's learning from historical training data as possible, thus achieving better model generalization and scenario adaptability.

[0038] Meanwhile, based on the construction of a small number of incremental data combined with historical core replay data, the training efficiency of tens of thousands of full historical data is greatly optimized, resulting in high training efficiency and fast iteration.

[0039] This avoids data privacy issues associated with large amounts of historical data, requiring only decryption processing of 100 core datasets. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0041] Example 1

[0042] A method for improving model performance using incremental learning includes the following steps:

[0043] Construct a historical core dataset based on historical training data;

[0044] The historical core dataset is anonymized to obtain the historical playback dataset;

[0045] Collect false alarm data during model operation and label the false alarm data to obtain a new scene training dataset;

[0046] The new scene training dataset is combined with the historical playback dataset to obtain an incremental training dataset, and the target model is incrementally trained based on a preset training strategy.

[0047] Based on the model parameters after the incremental training is completed, the optimized target model is output.

[0048] As one possible implementation, the steps of constructing a historical core dataset based on historical training data further include:

[0049] Obtain image data for each sample in the historical training dataset;

[0050] The image data is understood using a visual multimodal large model, and corresponding image text descriptions are generated based on preset prompts related to the target task.

[0051] The image text description is vectorized using a vision-language pre-trained model to obtain the feature vector of each image sample;

[0052] Clustering is performed based on the feature vectors to divide the historical training data into several clusters. The sample with the smallest distance to the feature vector of the cluster center is selected from each cluster as the representative sample of the cluster to form the historical core dataset.

[0053] As one possible implementation, the step of constructing a historical core dataset based on historical training data further includes:

[0054] While selecting representative samples based on feature vector clustering, samples are also selected from the historical training data and incorporated into the historical core dataset using at least one of the following methods:

[0055] A random sampling method is used to randomly select a preset number of samples from the historical training data;

[0056] The method for selecting difficult samples involves using existing models to predict historical training samples, sorting samples that are correctly predicted but have low confidence by category, and selecting a preset number of difficult samples from each category.

[0057] Based on the selection method of sample prediction uncertainty, different network types of models are used to predict the same sample for the same task. The difference in the prediction confidence of different models that predict the same category is used as the uncertainty index, and a preset number of samples are selected from the samples with larger differences.

[0058] As one possible implementation, the target tasks further include pedestrian detection tasks and / or vehicle detection tasks.

[0059] For the pedestrian detection task, the prompt words are used to guide the visual multimodal large model to output an image text description including the following:

[0060] The location, appearance, posture, behavior, and relationships between pedestrians or with the environment of each pedestrian;

[0061] For vehicle detection tasks, the cue words are used to guide the large visual multimodal model:

[0062] Determine whether a vehicle exists in the image. If a vehicle exists, output a text description for each vehicle, including at least one of the following: region location, vehicle type, color or paint scheme, orientation, degree of occlusion or truncation, abnormal state, environmental context, and potential detection challenges.

[0063] As one possible implementation, the preset training strategy further includes:

[0064] The parameters of the historical model, which are pre-trained based on the full historical training data, are used as the initialization parameters of the target model.

[0065] In the initial stage of training, the backbone network layer of the target model is frozen, and only the parameters of the neck network layer and the head network layer are updated. In the later stage of training, the backbone network layer is unfrozen to perform joint updates.

[0066] During each training iteration, a preset number of core samples are selected from the historical playback dataset and combined with samples from the new scene training dataset to form the current training batch.

[0067] As one possible implementation, the visual multimodal large model is further defined as at least one industry-specific visual multimodal large model, including any one or more of Tongyi Qianwen's Qwen2-VL or Qwen2.5-VL, OpenAI's GPT-4V, etc.

[0068] As one possible implementation, the vision-language pre-trained model is further selected from at least one of CLIP, BLIP-2, ViLT, and LLaVA.

[0069] As one possible implementation, the clustering is further performed using the K-means clustering algorithm, where the number of clusters is equal to the number of target historical core samples, and only one representative sample from each cluster is selected to be added to the historical core dataset.

[0070] As one possible implementation, the preset training strategy further includes:

[0071] During training, a learning rate that gradually decreases with the number of iterations is used to improve the model's convergence stability and reduce the forgetting of historical knowledge.

[0072] As a possible implementation, further, in each training iteration, in the step of selecting a preset number of core samples from the historical replay dataset and samples from the new scene training dataset to form the current training batch, the samples of the current training batch are composed of samples from the new scene training dataset and samples from the historical replay dataset in a preset ratio. The preset ratio is used to balance the adaptability of the new scene and the ability to retain historical knowledge during the training process.

[0073] Example 2

[0074] A method for improving model performance based on a small amount of incremental data, with the following brief steps:

[0075] Based on historical training data, a small number of core representation data were selected.

[0076] The core data is declassified and used as a thumbnail version of historical data, referred to here as playback data, which is then incorporated into subsequent model optimizations.

[0077] Collect online false positive data and use it as training data for new scenarios, then incorporate it into subsequent model optimization.

[0078] Adjust the conventional training strategy and optimize the training of the model to adapt to new scenarios while minimizing the forgetting of old knowledge.

[0079] By optimizing the training, the output model is obtained.

[0080] The specific method is described below:

[0081] Selection of historical core datasets

[0082] A serious drawback of incremental learning is catastrophic forgetting. To avoid catastrophic forgetting, all historical data is often added to the training, which leads to low training efficiency and data privacy issues. This invention does not use all historical data, but selects only a small amount of data from the full historical training data as the core representation of the dataset. This ensures that the model can recall historical knowledge from historical data and solves the problems of training efficiency and privacy. The method of selecting core data is related to the model's ability to retain historical performance.

[0083] Common methods for selecting core data, such as random sampling, selection based on difficult samples, and methods based on the uncertainty of sample prediction, can all be used as core data selection methods.

[0084] Based on random sampling, that is, randomly selecting a number of samples, for example, 100 samples can be randomly selected as representative of historical data.

[0085] Based on difficult samples, existing pre-trained prediction models are used. For example, for classification tasks, the model is the sample classification model (such as ResNet50, VGGNet, etc., which are common pre-trained classification models), and for detection tasks, the model is the sample detection model (such as YOLOv1~V11, etc., YOLO series). The training samples are predicted and the predicted class and confidence score are output. The prediction confidence scores of the samples are sorted according to different classes. A certain number of samples with correct class predictions but low prediction confidence scores are selected as the difficult sample set for that class. The difficult sample sets of all classes are integrated to construct the core difficult sample set of historical data. This difficult set is used as the historical core dataset.

[0086] The method based on sample prediction uncertainty is similar to that based on difficult samples. However, it uses different network types of classification or detection models for the same task and makes predictions on the same sample. The confidence scores of different models predicting the same class are compared, and all sample differences are sorted by class. The larger the difference, the greater the uncertainty of the sample. A certain number of samples with large differences are selected to construct the core uncertain sample set of the class. The uncertain sample sets of all classes are integrated to construct the core uncertain sample set of historical data, which serves as the historical core dataset.

[0087] This invention proposes another method for selecting core datasets based on the representational features of large models.

[0088] A. Based on industry-specific visual multimodal large models, such as Alibaba's Tongyi Qianwen's qwen2vl / qwen2.5vl and OpenAI's GPT-4V, images are understood and image descriptions are generated. Appropriate prompt words can be set for different tasks. For example, for pedestrian detection tasks, the prompt words can be designed as follows:

[0089] Please provide a detailed description of this image, focusing on all pedestrians. For each pedestrian, please indicate their location (using relative coordinates or image area descriptions), physical characteristics (e.g., gender, age, clothing, color, whether wearing a hat / backpack), posture (standing, walking, running, cycling, etc.), behavior (making a phone call, looking at a mobile phone, pushing a stroller, etc.), and their relationship to the environment (e.g., whether they are obstructed, in a crowd, or near vehicles). If no pedestrians are visible in the image, please clearly state "No pedestrians detected in the image."

[0090] For vehicle inspection, the prompt message can be designed as follows:

[0091] "First, describe whether any type of vehicle (including sedans, SUVs, buses, trucks, vans, tricycles, motorcycles, bicycles, etc.) exists in the scene. If none are found, output: No vehicle detected in the image, and end the response. If vehicles are found, describe each vehicle using the following fields, separated by "|", one vehicle per line: Vehicle number | Region location | Vehicle type | Color / paint | Orientation | Occlusion or truncation level | Abnormal state | Environmental context | Potential detection challenges."

[0092] By asking questions of a large model, a detailed description of the image is generated for the specific task.

[0093] B. Vectorize the image descriptions generated for all images in the training set into text.

[0094] Text vectorization is based on Vision-Language pre-trained models, such as CLIP, BLIP-2, ViLT, and LLaVA. The differences between these models lie only in the number of layers, attention masks, and positional encoding. All of them can convert image descriptions into feature vectors for image retrieval and text-image retrieval.

[0095] C. Perform similarity clustering based on the feature vectors of all images in the training set.

[0096] Based on the feature vectors generated in the previous step, feature similarity clustering is performed. An effective way is to use k-means clustering. k-means is a centroid-based clustering algorithm. When it performs a task, it makes decisions based on the centroid of each cluster and the distance between each data point and the cluster center, automatically classifying k groups. Here, k-means is used to automatically divide the data into k groups, where k is the number of core set images to be output, which is defined by the user. When the k groups are formed, the sample whose feature is closest to the center of the k groups is selected as the core data sample. This ensures that the feature distance is closest within the k groups and farthest outside the groups. Here, k is set to 100.

[0097] D. Based on the above clustering results, output the core dataset of historical data.

[0098] In this way, the selected core sample set, after eliminating similar sample sets, has rich scene coverage and provides a more accurate representation of historical sample sets.

[0099] 2. The core set is de-encrypted and used as a thumbnail of historical data, referred to here as replay data, which is then incorporated into subsequent model optimizations.

[0100] To reduce data leakage, data is declassified. This declassification process mainly involves blurring out private information such as faces, license plates, or addresses in images.

[0101] 3. Collect online false positive data and use it as training data for new scenarios, incorporating it into subsequent model optimization.

[0102] Collect online false alarm data and manually label the data as a data source for subsequent model adaptation to new scenarios. At the same time, alarm data can also be used directly to assist the model in understanding the scenario. The more data collected for new scenarios, the better the adaptability to new scenarios. It is recommended to collect more than 1,000 data points.

[0103] 4. Adjust the conventional training strategy and optimize the training of the model to adapt to new scenarios while minimizing the forgetting of old knowledge.

[0104] After collecting false positive data for new scenarios and historical core datasets, we can start optimizing and training the model. Depending on the task type, we can choose a classification network or a detection network. Here, we will take the optimization of the detection model YOLO as an example.

[0105] Conventional training strategies typically initialize parameters based on publicly available pre-trained network models, optimize all network parameters according to the input dataset, and output the model after meeting the required number of iterations or loss conditions. This training method results in poor model training performance with limited sample data, poor adaptability to different scenarios, and a high rate of forgetting historical data. This invention adjusts the training strategy to better adapt to new scenarios while minimizing the forgetting of old knowledge.

[0106] A. During the model parameter initialization phase, pre-trained model parameters based on all historical data are used as initialization parameters. This provides prior knowledge of historical data, making convergence easier and allowing the model to retain richer historical knowledge. This model can be placed as built-in pre-set parameters during network training and directly loaded in future training for new scenarios.

[0107] B. The YOLO network layer consists of backbone, neck, and head layers. At the beginning of training, the backbone of the network is frozen and no parameters are updated. Only the neck and head are updated. The backbone is unfrozen and updated in the later stages of the network iteration. The synchronous learning rate is updated by gradually decreasing as the number of iterations increases. In this way, the model can retain as much knowledge as possible from a large amount of historical data and adapt to new scenarios.

[0108] C. For the training process, all data is generally added evenly to the optimization according to the set batch size. In order to retain more learning from historical core data, several core data will be selected and added to the training in each iteration. For example, if the batch size is set to 16, 4 to 5 core data can be added in each iteration. This training method can also increase the retention of historical knowledge.

[0109] 5. Output the model through optimized training.

[0110] The above description is an embodiment of the present invention. For those skilled in the art, any equivalent changes, modifications, substitutions and variations made in accordance with the teachings of the present invention without departing from the principles and spirit of the present invention should be covered by the present invention.

Claims

1. A method for improving model performance using incremental learning, characterized in that, Includes the following steps: Construct a historical core dataset based on historical training data; The historical core dataset is anonymized to obtain the historical playback dataset; Collect false alarm data during model operation and label the false alarm data to obtain a new scene training dataset; The new scene training dataset is combined with the historical playback dataset to obtain an incremental training dataset, and the target model is incrementally trained based on a preset training strategy. Based on the model parameters after the incremental training is completed, the optimized target model is output.

2. The method for improving model performance using incremental learning as described in claim 1, characterized in that: The steps for constructing the historical core dataset based on historical training data include: Obtain image data for each sample in the historical training dataset; The image data is understood using a visual multimodal large model, and corresponding image text descriptions are generated based on preset prompts related to the target task. The image text description is vectorized using a vision-language pre-trained model to obtain the feature vector of each image sample; Clustering is performed based on the feature vectors to divide the historical training data into several clusters. The sample with the smallest distance to the feature vector of the cluster center is selected from each cluster as the representative sample of the cluster to form the historical core dataset.

3. A method for improving model performance using incremental learning as described in claim 1 or 2, characterized in that: The step of constructing the historical core dataset based on historical training data also includes: While selecting representative samples based on feature vector clustering, samples are also selected from the historical training data and incorporated into the historical core dataset using at least one of the following methods: A random sampling method is used to randomly select a preset number of samples from the historical training data; The method for selecting difficult samples involves using existing models to predict historical training samples, sorting samples that are correctly predicted but have low confidence by category, and selecting a preset number of difficult samples from each category. Based on the selection method of sample prediction uncertainty, different network types of models are used to predict the same sample for the same task. The difference in the prediction confidence of different models that predict the same category is used as the uncertainty index, and a preset number of samples are selected from the samples with larger differences.

4. A method for improving model performance using incremental learning as described in claim 1 or 2, characterized in that: The target tasks include pedestrian detection and / or vehicle detection. For the pedestrian detection task, the prompt words are used to guide the visual multimodal large model to output an image text description including the following: The location, appearance, posture, behavior, and relationships between pedestrians or with the environment of each pedestrian; For vehicle detection tasks, the cue words are used to guide the large visual multimodal model: Determine whether a vehicle exists in the image. If a vehicle exists, output a text description for each vehicle, including at least one of the following: region location, vehicle type, color or paint scheme, orientation, degree of occlusion or truncation, abnormal state, environmental context, and potential detection challenges.

5. The method for improving model performance using incremental learning as described in claim 1, characterized in that: The preset training strategy includes: The parameters of the historical model, which are pre-trained based on the full historical training data, are used as the initialization parameters of the target model. In the initial stage of training, the backbone network layer of the target model is frozen, and only the parameters of the neck network layer and the head network layer are updated. In the later stage of training, the backbone network layer is unfrozen to perform joint updates. During each training iteration, a preset number of core samples are selected from the historical playback dataset and combined with samples from the new scene training dataset to form the current training batch.

6. The method for improving model performance using incremental learning as described in claim 1, characterized in that: The visual multimodal large model is at least one industry-specific visual multimodal large model, including any one or more of Tongyi Qianwen's Qwen2-VL or Qwen2.5-VL, OpenAI's GPT-4V, etc.

7. The method for improving model performance using incremental learning as described in claim 1, characterized in that: The vision-language pre-trained model is at least one of CLIP, BLIP-2, ViLT, and LLaVA.

8. The method for improving model performance using incremental learning as described in claim 1, characterized in that: The clustering is performed using the K-means clustering algorithm. The number of clusters is equal to the number of target historical core samples, and only one representative sample is selected from each cluster to be added to the historical core dataset.

9. A method for improving model performance using incremental learning as described in claim 5, characterized in that: The preset training strategy also includes: During training, a learning rate that gradually decreases with the number of iterations is used to improve the model's convergence stability and reduce the forgetting of historical knowledge.

10. A method for improving model performance using incremental learning as described in claim 5, characterized in that: In the step of selecting a preset number of core samples from the historical replay dataset and samples from the new scene training dataset to form the current training batch during each training iteration, the samples in the current training batch are composed of samples from the new scene training dataset and samples from the historical replay dataset in a preset ratio. The preset ratio is used to balance the adaptability to the new scene and the ability to retain historical knowledge during the training process.