A method and system for accelerating active learning training based on database indexing technology

Abstract

The application discloses a kind of active learning training acceleration method and system based on database index technology.Method includes: training sample is input into sample feature extraction module, and output feature vector;Input active learning evaluation module, and output sample score, call high-performance index module sorting;High-performance index module saves and maintains order relationship to sample score;Integrative active learning algorithm module calls high-performance index module pre-screening, uses active learning algorithm to select training sample, and input into the deep learning model of training;Repeat until training is completed.The application combines active learning and efficient data index structure, speeds up the active learning training process, improves the efficiency and performance of model training.Optimize the storage and indexing method of data, reduce the complexity of data operation, help to improve the overall efficiency of the training process, can be based on efficient database index structure, and light active learning algorithm, accelerate the training process.

Description

Technical Field

[0001] This invention relates to an active learning training acceleration method, which relates to the field of artificial intelligence (AI) database indexing technology, and specifically to an active learning training acceleration method and system based on database indexing technology. Background Technology

[0002] In recent years, deep neural networks have made remarkable progress in various fields and have become a core technology for many tasks. However, training these deep neural networks typically requires a large amount of high-quality labeled data to achieve optimal performance. For example, ChatGPT requires a corpus of tens of millions of sentences for its complete training. Although some large datasets exist for training deep learning models, their quality and quantity still fall short of the requirements for training deep neural network models. Therefore, active learning, a technique that effectively selects and labels the most valuable samples, has been widely applied in deep learning model training to reduce labeling costs and improve model performance.

[0003] Active learning, as an effective learning paradigm, reduces annotation costs and improves model performance by selectively labeling the most informative samples. In active learning, the key to sample selection is the ability to quickly and efficiently find the most effective samples for training the current model. In recent years, researchers have proposed many methods and techniques applicable to active learning to improve the efficiency and performance of sample selection.

[0004] While research on active learning has yielded some results, traditional sample selection methods face several challenges when dealing with large-scale datasets.

[0005] 1. The problem of repetitive training: In traditional deep learning model training, due to the existence of many similar data samples, the model may undergo a large amount of repetitive training, making it difficult to effectively retrieve samples that are more useful to the model. Therefore, it is necessary to develop more intelligent and efficient sample selection strategies to optimize the training efficiency and performance of the model.

[0006] 2. Inefficient data scanning and repositioning: Traditional active learning sample selection methods may face inefficient data scanning and repositioning issues on large-scale datasets. This leads to decreased training efficiency on large datasets and increases computational costs during training.

[0007] 3. Maintenance Overhead: When a data index requires extensive updates, such as adding or deleting data, involving data movement and reordering, this incurs significant maintenance overhead. Therefore, more efficient data indexing and update methods are needed to reduce maintenance costs and maintain dataset stability.

[0008] Traditional deep learning model training processes often use a large number of similar data samples, leading to repetitive training and hindering optimal performance. Furthermore, traditional active learning sample selection methods may suffer from inefficient data scanning and repositioning issues on large datasets. This results in decreased training efficiency on large datasets and increases computational costs during training. Summary of the Invention

[0009] To address the problems existing in the background technology, the present invention provides an active learning training acceleration method and system based on database indexing technology.

[0010] The technical solution adopted in this invention is:

[0011] I. A method for accelerating active learning training based on database indexing technology, comprising:

[0012] Step 1) Establish an active learning training acceleration model, which includes a sample feature extraction module, an active learning evaluation module, a high-performance indexing module, and an integrated active learning algorithm module.

[0013] Step 2) Input each untrained sample and trained sample into the sample feature extraction module, and the sample feature extraction module outputs the feature vectors of each untrained sample and trained sample.

[0014] Step 3) Input the feature vectors of each untrained sample and trained sample into the active learning evaluation module. The active learning evaluation module outputs the sample score of each untrained sample. The high-performance indexing module is called to sort the sample scores of each untrained sample from high to low.

[0015] Step 4) The high-performance indexing module saves the sample scores of each untrained sample, and at the same time maintains the sorting order of the sample scores of each untrained sample.

[0016] Step 5) The integrated active learning algorithm module pre-screens each untrained sample by calling the high-performance indexing module to obtain several boundary samples. Then, the integrated active learning algorithm module uses the active learning algorithm to select several training samples from each boundary sample and inputs each training sample into the deep learning model to be trained for training.

[0017] Step 6) Repeat steps 1)-5) until training is complete.

[0018] In step 2), the sample feature extraction module is specifically a self-supervised DINO (DETR with Improved deNoising anchOr boxes) model. Before extracting sample features, the training samples need to undergo standard preprocessing.

[0019] In step 3), the feature vectors of each untrained sample and trained sample are input into the active learning evaluation module. The active learning evaluation module first uses the k-Nearest Neighbors (KNN) algorithm to search and select the feature vectors of the nearest untrained samples of each trained sample's feature vector. Then, the frequency of the selected feature vector of each untrained sample is obtained as its respective sample score.

[0020] During the initialization phase of the active learning evaluation module, the initial sample score for the feature vector of each untrained sample is 0. x ′} init =0, x′∈D unlabeld Subsequently, in each training round, the sample score of the feature vector of each untrained sample is iteratively updated, as follows:

[0021]

[0022] in, and x represents the sample score of the feature vector of the untrained sample in the current iteration update and the previous iteration update, respectively; i and x′ i D represents the feature vector of a trained sample and the feature vector of its nearest neighbor untrained sample selected by the k-nearest neighbor (KNN) algorithm, respectively; labeled This represents the set of feature vectors of the trained samples.

[0023] In step 4), the high-performance indexing module includes a semi-ordered index structure and a status information record table. The semi-ordered index structure includes several data blocks, and the data blocks maintain an ordered relationship. Each data block stores several data items, and the data items maintain an unordered relationship. Each data item includes an untrained sample and its sample score. When inserting data into the index, the system will traverse each data block sequentially from the beginning of the index, find the first data block whose data is located within the storage range of that block, and then store the data in the free space within the block.

[0024] The status information record table records the status information of each data item in each data block, that is, the update status of the feature vector of each untrained sample. The update status includes the sample score of the feature vector of the untrained sample being updated, not being updated, and already updated.

[0025] The high-performance indexing module maintains sample score rankings based on an updated and optimized database index structure, enabling other modules to quickly retrieve the best samples.

[0026] In step 4), the high-performance indexing module maintains the sorting order of the sample scores for each untrained sample. Specifically, for the status information of each data item in each data block, when the status information of a data item is in the "being updated" state, there is no need to traverse the index to update the sample scores in the data item; only the frequency increase of the current update needs to be updated in the status information record table. When the status information of a data item is in the "not being updated" state, the status information of the data item is adjusted to the "being updated" state, and the index is traversed to update the sample scores in the data item, adjusting the position of the sample in the index.

[0027] In step 5), the integrated active learning algorithm module pre-screens each untrained sample to obtain several boundary samples by calling the high-performance indexing module. Specifically, under the condition of maximizing the screening objective, several untrained samples with higher scores are pre-screened as boundary samples. The maximizing screening objective is as follows:

[0028]

[0029] Among them, D selected represents the pre-selected set of boundary samples, which includes several boundary samples; x represents untrained samples; I() and C() represent the information function and contribution function, respectively. I() is obtained using the confidence level obtained from model inference, and C() is calculated based on the correlation between untrained samples and trained samples.

[0030] Given a training dataset Where x i It is the input sample, y i It consists of the labels corresponding to the input samples and a model f(x; θ), where θ are the parameters of the trained model, and x are the training data input into the model. The goal of the module is to obtain data from D. unlabeld Maximize the selection of sample set The total amount of information and the total contribution.

[0031] In step 5), the active learning algorithm is specifically the minimum confidence algorithm or the highest entropy algorithm, etc.

[0032] The integrated active learning algorithm module uses existing active learning algorithms in D selected The next step is to select the most valuable samples. In this module, the indexing module is used to select the batch of samples with the highest scores as D. selectedThen, on this batch of samples, existing active learning algorithms are used. selected It contains less data than D unlabel Therefore, this method reduces screening costs and increases the speed of the training process.

[0033] The samples used in training within the integrated active learning algorithm module will call the interface of the active learning evaluation module to update the scores of unlabeled samples. Furthermore, this module will also adjust the order of samples in the high-performance indexing module through the active learning evaluation module.

[0034] II. An active learning training acceleration model based on an active learning training acceleration method:

[0035] The model includes a sample feature extraction module for extracting feature vectors from each untrained and trained input sample.

[0036] The model includes an active learning evaluation module that uses the feature vectors output by the sample feature extraction module to actively learn and evaluate the samples to obtain sample scores, and calls the high-performance indexing module to adjust the order of the samples.

[0037] The model includes a high-performance indexing module for saving and updating sample scores and for sorting.

[0038] The model includes an integrated active learning algorithm module that calls a high-performance indexing module for pre-screening and then performs final screening to obtain the training samples.

[0039] The beneficial effects of this invention are:

[0040] This invention combines active learning with an efficient data indexing structure to accelerate the active learning training process, proposing a novel active learning training strategy that can improve the efficiency and performance of model training. Simultaneously, it optimizes data storage and indexing methods, reducing the complexity of data operations, which helps lower maintenance overhead and improve the overall efficiency of the training process. Based on an efficient database indexing structure and a lightweight active learning algorithm, it accelerates the training process. Attached Figure Description

[0041] Figure 1 This is a flowchart of the implementation steps of the present invention;

[0042] Figure 2 This is a design diagram of the high-performance index module of the present invention. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] like Figure 1 As shown, the active learning training acceleration method based on database indexing technology includes:

[0045] Step 1) Establish an active learning training acceleration model, which includes a sample feature extraction module, an active learning evaluation module, a high-performance indexing module, and an integrated active learning algorithm module.

[0046] Step 2) Input each untrained sample and trained sample into the sample feature extraction module, and the sample feature extraction module outputs the feature vectors of each untrained sample and trained sample.

[0047] In step 2), the sample feature extraction module is specifically a self-supervised DINO (DETR with Improved deNoisinganchOr boxes) model. Before performing sample feature extraction, the training samples need to undergo a standardized preprocessing operation.

[0048] Step 3) Input the feature vectors of each untrained sample and trained sample into the active learning evaluation module. The active learning evaluation module outputs the sample score of each untrained sample and sorts the sample scores of each untrained sample from high to low by calling the high-performance indexing module.

[0049] In step 3), the feature vectors of each untrained sample and trained sample are input into the active learning evaluation module. The active learning evaluation module first uses the k-Nearest Neighbors (KNN) algorithm to search and select the feature vectors of the nearest untrained samples of each trained sample's feature vector. Then, the frequency of the selection of the feature vector of each untrained sample is obtained as its respective sample score.

[0050] During the initialization phase of the active learning evaluation module, the initial sample score for the feature vector of each untrained sample is 0. x ′} init =0, x′∈D unlabeld Subsequently, in each training round, the sample score of the feature vector of each untrained sample is iteratively updated, as follows:

[0051]

[0052] in, and x represents the sample score of the feature vector of the untrained sample in the current iteration update and the previous iteration update, respectively; i and x i ′ represent the feature vectors of the trained samples and the feature vectors of their nearest untrained neighbors selected by the k-nearest neighbor (KNN) algorithm, respectively; D labeled This represents the set of feature vectors of the trained samples.

[0053] Step 4) The high-performance indexing module saves the sample scores of each untrained sample, and at the same time maintains the sorting order of the sample scores of each untrained sample.

[0054] In step 4), the high-performance indexing module includes a semi-ordered index structure and a status information record table. The semi-ordered index structure includes several data blocks, and the data blocks maintain an ordered relationship. Each data block stores several data items, and the data items maintain an unordered relationship. Each data item includes an untrained sample and its sample score. When inserting data into the index, it will traverse each data block sequentially from the head of the index, find the first data block whose data is located within the storage range of that block, and then store the data in the free space within the block.

[0055] The status information record table records the status information of each data item in each data block, that is, the update status of the feature vector of each untrained sample. The update status includes the sample score of the feature vector of the untrained sample being updated, not being updated, and already updated.

[0056] The high-performance indexing module maintains sample score rankings based on an updated and optimized database index structure, enabling other modules to quickly retrieve the best samples.

[0057] In step 4), the high-performance indexing module maintains the sorting order of the sample scores for each untrained sample. For the status information of each data item in each data block, specifically, when the status information of a data item is in the "being updated" state, there is no need to traverse the index to update the sample scores in the data item; only the frequency increase of the current update needs to be updated in the status information record table. When the status information of a data item is in the "not being updated" state, the status information of the data item is adjusted to the "being updated" state, and the index is traversed to update the sample scores in the data item, adjusting the position of the sample in the index.

[0058] Step 5) The integrated active learning algorithm module pre-screens each untrained sample by calling the high-performance indexing module to obtain several boundary samples. Then, the integrated active learning algorithm module uses the active learning algorithm to select several training samples from each boundary sample and inputs each training sample into the deep learning model to be trained for training.

[0059] In step 5), the integrated active learning algorithm module pre-screens each untrained sample to obtain several boundary samples by calling the high-performance indexing module. Specifically, under the condition of maximizing the screening objective, it pre-screens several untrained samples with higher scores as boundary samples. The maximizing screening objective is as follows:

[0060]

[0061] Among them, D selected represents the pre-selected set of boundary samples, which includes several boundary samples; x represents untrained samples; I() and C() represent the information function and contribution function, respectively. I() is obtained using the confidence level obtained from model inference, and C() is calculated based on the correlation between untrained samples and trained samples.

[0062] Given a training dataset Where x i It is the input sample, y i It consists of the labels corresponding to the input samples and a model f(x; θ), where θ are the parameters of the trained model, and x are the training data input into the model. The goal of the module is to obtain data from D. unlabeld Maximize the selection of sample set The total amount of information and the total contribution.

[0063] In step 5), the active learning algorithm is specifically the minimum confidence algorithm or the highest entropy algorithm, etc.

[0064] The integrated active learning algorithm module uses existing active learning algorithms in D selected The next step is to select the most valuable samples. In this module, the indexing module is used to select the batch of samples with the highest scores as D. selected Then, on this batch of samples, existing active learning algorithms are used. selected It contains less data than D unlabel Therefore, this method reduces screening costs and increases the speed of the training process.

[0065] The samples used in training within the integrated active learning algorithm module will call the interface of the active learning evaluation module to update the scores of unlabeled samples. Furthermore, this module will also adjust the order of samples in the high-performance indexing module through the active learning evaluation module.

[0066] Step 6) Repeat steps 1)-5) until training is complete.

[0067] The active learning training acceleration model of the present invention includes a sample feature extraction module for extracting features from each untrained sample and trained sample input sample to obtain feature vectors; an active learning evaluation module for obtaining sample scores by performing active learning evaluation based on the feature vectors output by the sample feature extraction module and calling a high-performance index module to adjust the order of samples; a high-performance index module for saving and updating sample scores and sorting them; and an integrated active learning algorithm module for calling the high-performance index module for pre-screening and then for final screening to obtain samples to be trained.

[0068] like Figure 2As shown, the high-performance indexing module consists of a semi-ordered index structure and a status information record table. The former is a linked structure, where each data node is a block that can hold multiple data items. The data items within each block are unordered, but the blocks themselves are ordered. The status information record table records the update status of each data item. Threads within the high-performance indexing module access and update the data nodes.

[0069] The untrained sample x and the trained sample pool D of this invention train Feature vector representations are obtained through a pre-trained sample representation model. Then, based on these feature vectors, the KNN nearest neighbors of each trained sample are calculated. Simultaneously, a sample score is calculated based on the frequency of untrained samples x as nearest neighbors. This process maintains the score of each sample using a semi-ordered index structure. After obtaining the scores of all untrained samples, the sample x′ with the highest score is sequentially read from the index as the boundary sample and combined into D. selected Finally, in D selected Then, using existing active learning algorithms, select a batch of samples that yield the greatest benefit to the model, and then label and use them in model training.

[0070] For active learning scenarios, this invention proposes an active learning algorithm that can perceive the inter-class boundaries of the model on a specific dataset, calculate the training value of each unlabeled sample based on the boundary, and then select the batch of samples with the highest value to apply the active learning algorithm, thereby intelligently selecting the optimal samples, ultimately reducing the number of samples required for training and accelerating the overall active learning training process.

Claims

1. A method for accelerating active learning training based on database indexing technology, characterized in that, The method comprises the following steps: Step 1) establishing an active learning training acceleration model, the active learning training acceleration model comprising a sample feature extraction module, an active learning evaluation module, a high-performance index module and an integrated active learning algorithm module; Step 2) inputting each untrained sample and trained sample into the sample feature extraction module, and outputting the feature vectors of each untrained sample and trained sample by the sample feature extraction module; Step 3) inputting the feature vectors of each untrained sample and trained sample into the active learning evaluation module, and outputting the sample scores of each untrained sample by the active learning evaluation module, and sorting the sample scores of each untrained sample from high to low by calling the high-performance index module; Step 4) saving the sample scores of each untrained sample by the high-performance index module, and maintaining the order relationship of the sample scores of each untrained sample by the high-performance index module; Step 5) pre-screening each untrained sample to obtain a plurality of boundary samples by the integrated active learning algorithm module calling the high-performance index module, and then selecting a plurality of training samples from the boundary samples by the integrated active learning algorithm module using an active learning algorithm, and inputting each training sample into a deep learning model to be trained; Step 6) repeating steps 1) to 5) until the training is completed; In step 4), the high-performance index module comprises a semi-ordered index structure and a state information record table, the semi-ordered index structure comprising a plurality of data blocks, the data blocks maintaining an ordered relationship, each data block saving a plurality of data items, the data items maintaining an unordered relationship, and each data item comprising an untrained sample and a sample score of the untrained sample; The state information record table records the state information of each data item in each data block, i.e., the update state of the feature vector of each untrained sample, the update state comprising an untrained sample feature vector sample score being updated, not being updated and having been updated; In step 5), the integrated active learning algorithm module pre-screens each untrained sample to obtain a plurality of boundary samples by calling the high-performance index module, specifically pre-screening a plurality of untrained samples with higher sample scores from each untrained sample as boundary samples under the condition of meeting the maximum screening target, and the maximum screening target being specifically as follows: Wherein, Indicates a pre-screened boundary sample set, and the boundary sample set includes a plurality of boundary samples; Indicates an untrained sample; And Indicates an information amount function and a contribution function, respectively.

2. The active learning training acceleration method based on database indexing techniques of claim 1, wherein: In step 2), the sample feature extraction module is specifically a self-supervised DINO model.

3. The database index technology based active learning training acceleration method of claim 1, wherein: In step 3), the feature vectors of each untrained sample and trained sample are input into the active learning evaluation module, the active learning evaluation module first uses a k-nearest neighbor (KNN) algorithm to search and select the feature vectors of the nearest neighbor untrained samples of each trained sample, and then obtains the frequency of each untrained sample feature vector being selected as the sample score of each untrained sample; In the initialization stage of the active learning evaluation module, the initial sample score of each untrained sample feature vector is 0, and the sample score of each untrained sample feature vector is iteratively updated in each training round, and the specific process is as follows: wherein, and respectively represent the sample score of the feature vector of the untrained sample of the current iteration update and the last iteration update; and respectively represent the feature vector of the trained sample and the feature vector of the nearest neighbor untrained sample selected by the k nearest neighbor (KNN) algorithm; represents the set of the feature vector of the trained sample.

4. The database index technology based active learning training acceleration method of claim 1, wherein: In the step 4), the high-performance index module maintains the order relationship of the sample scores of the respective untrained samples, and for the state information of each data item in each data block, when the state information of the data item is in the state of being updated, the sample score in the data item does not need to be updated by traversing the index, and only the increased frequency of the current update needs to be updated to the state information record table; when the state information of the data item is in the state of not being updated, the state of the state information of the data item is adjusted to the state of being updated, and the sample score in the data item is updated by traversing the index.

5. The database index technology based active learning training acceleration method of claim 1, wherein: In the step 5), the active learning algorithm is specifically a minimum confidence algorithm or an entropy highest algorithm.

6. An active learning training acceleration model based on the active learning training acceleration method of any one of claims 1-5. The sample feature extraction module is configured to extract features of the respective untrained samples and the trained samples to obtain feature vectors. The active learning evaluation module is configured to obtain sample scores based on the feature vectors output by the sample feature extraction module and to call the high-performance index module to adjust the order relationship of the samples. The high-performance index module is configured to save and update the sample scores and to sort the sample scores. The integrated active learning algorithm module is configured to call the high-performance index module to perform pre-screening and then to perform final screening to obtain the to-be-trained samples.

Images