Model data reinforcement optimization method and device based on information gain optimization
By optimizing data batches using information entropy theory and greedy methods, the performance of autonomous driving perception models in complex testing scenarios is improved, solving the problem of model performance degradation in existing technologies and achieving efficient data hardening.
Patent Information
- Application Number
- CN202510889205.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-28
AI Technical Summary
Existing autonomous driving perception models experience performance degradation when faced with discrepancies between actual data distribution and training data distribution. Current hardening techniques have failed to effectively improve model performance in complex testing scenarios, and the training process is either complex or yields unsatisfactory results.
By establishing an information gain metric based on information entropy theory, designing a greedy method to construct data batches, optimizing the performance of data batches in the candidate data pool on the test environment, and introducing mutual information of uncertain prediction results as the basis for computation, efficient data hardening is achieved.
Under the same sample selection restrictions, the performance of the autonomous driving perception model in complex test scenarios is significantly improved, the performance degradation problem caused by data distribution differences in existing technologies is solved, and redundant information and training complexity are avoided.
Smart Images
Figure CN120852902A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of autonomous driving and machine learning, and specifically relates to a method and apparatus for strengthening and optimizing model data based on information gain optimization. Background Technology
[0002] Autonomous driving heavily relies on artificial intelligence technologies, including perception, prediction, and planning / control. Perception models, used to understand the vehicle's surrounding environment, are a typical example. However, in practical deployments, autonomous driving perception models face the challenges of complex, dynamic, and diverse testing scenarios. Furthermore, when there is a significant difference between the actual data distribution and the training data distribution, the performance of perception models often deteriorates severely, threatening and challenging the safe and reliable use of the models. Therefore, a series of hardening techniques that improve models or augment data have been proposed. Existing technologies mainly focus on the following aspects:
[0003] 1) Model learning techniques based on invariant representation extraction.
[0004] Invariant representation extraction aims to learn cross-domain invariant representations using labels from different data domains, and then perform model inference and prediction based on these invariant representations. The most representative technique, invariant risk minimization, introduces constraints during model learning to ensure that the optimal inference network pointed to by the representations of the intermediate layers is shared across different data domains, thus making the intermediate layer representations invariant. Other techniques achieve invariant representation extraction by minimizing the distribution distance of representations between different data domains. However, these techniques complicate the training process of intelligent models, making training more difficult. Furthermore, practice shows that models trained using these techniques do not perform ideally. Therefore, data-level model optimization solutions remain urgently needed.
[0005] 2) Optimization technique using a split-bar.
[0006] Split-Brow optimization techniques hypothesize a set of potential actual distributions, forming a potential target distribution set. Robustness of the intelligent model is achieved by optimizing the worst-case performance of the model within this set. By varying the method of constructing the potential target distribution set, a series of split-Brow optimization methods have emerged. For example, the potential target distribution set can be defined as the set of distributions whose distance from the training set distribution is less than a threshold, where the distance can be f-divergence or bulldozer movement distance, etc. Alternatively, the potential target distribution set can be defined as the distribution obtained by adjusting the proportions of different subgroups. However, in order to ensure that the potential target distribution set includes the target distribution, these techniques often define the potential target distribution set too large, causing the training of the intelligent model to be influenced by irrelevant distributions, thus negatively impacting the performance of the trained model.
[0007] 3) Data augmentation-based reinforcement and optimization techniques.
[0008] Data augmentation-based hardening and optimization techniques aim to select data with the potential to improve model performance from a candidate data pool. To this end, different data selection techniques employ various data evaluation criteria, including the current model's predictive uncertainty, the influence of model parameters, and the representativeness of the data distribution. However, these techniques fail to accurately calculate the impact of data on the model's predictions on the target test samples, and they do not consider the migration and changes in data distribution relative to the candidate data pool in actual model deployment scenarios. This prevents these techniques from achieving the desired level of optimization for intelligent models.
[0009] In the field of autonomous driving, model learning techniques based on invariant representation extraction require labels in the data domain. However, in autonomous driving applications, these labels are often difficult to obtain in advance, limiting the practical application of this type of method. Furthermore, this approach makes model training unstable and difficult to converge. Additionally, the performance of models trained using this technique is not ideal.
[0010] To ensure that the potential target distribution set encompasses the target distribution, the Bruker optimization method often defines the range of the potential target distribution set to be quite large. This makes the training of the intelligent model susceptible to the influence of irrelevant distributions, especially the impact of noisy samples. This causes the trained model to overemphasize data samples irrelevant to the application. In vision tasks such as autonomous driving perception, data is often high-dimensional and possesses an inherent data structure. Defining the potential data distribution using a simple distance function will generate many data distributions and noisy samples that are meaningless in real-world autonomous driving scenarios, thus negatively impacting the learning of the perception model.
[0011] Data augmentation-based hardening optimization methods focus on the properties of the data samples themselves and their relationship with model parameters, but neglect the impact of data samples on the model's performance under autonomous driving perception test samples, and also ignore the distributional differences between the candidate sample data pool and the perception test sample distribution. These factors lead to unsatisfactory performance of data augmentation-based hardening optimization techniques in real-world testing environments. For example, Bayesian active learning techniques based on diversified gradient representations construct gradient vectors for each sample and use clustering algorithms to find the cluster centers of the gradient vectors, using these as the selected samples. This technique does not consider the information gain of the prediction results for the target test samples, resulting in unsatisfactory performance hardening of the model under the target test scenario. The above schemes ignore the mismatch between the data candidate pool and the target test sample distribution, making the direct selection of representative samples from the data candidate pool ineffective in hardening the performance of the target test samples.
[0012] In addition, some techniques, such as the expected prediction information gain optimization technique, only calculate the information gain for a single data point. However, the reinforcement scheme of collecting a single data point in each round results in a very long process cycle, which is not conducive to application in scenarios involving massive training data, such as autonomous driving. When directly extended to data batch scenarios, the diversity between data sample points in the batch is ignored, leading to the problem of redundant information in the acquired data. The intelligent model trained after being put into the training set with data batches containing redundant information performs poorly in the target test scenario. Summary of the Invention
[0013] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a model data hardening optimization method and apparatus based on information gain optimization. This invention establishes an information metric criterion, based on information entropy theory, to characterize the degree of performance improvement of various data batches in the candidate data pool on the perception model under test conditions. Through reasoning and analysis of information gain measurement, an efficient greedy method is designed to construct data batches, thereby completing the data hardening of model performance. This invention has the advantage of maximizing the performance improvement of the intelligent model under the same sample selection constraints, providing important technical support for the application of autonomous driving perception models in complex and unseen test scenarios.
[0014] A first aspect of this invention proposes a model data hardening optimization method based on information gain optimization, comprising:
[0015] The data points that can be used as input data for the intelligent model are formed into a candidate data pool, and the total number of rounds to select data points from the candidate data pool is set.
[0016] When each data point selection round arrives, the data point that maximizes the improvement of the information gain metric estimation result is searched from the current candidate data pool and added to the data batch saved in the previous round to obtain the current data batch; when the total number of data points in the current data batch is equal to the total number of rounds, the final data batch is generated.
[0017] Each data point in the final data batch is labeled with a corresponding tag to form a new training sample. Then, all the new training samples are added to the original training dataset of the intelligent model to complete the data hardening of the training dataset.
[0018] In one specific embodiment of the present invention, it further includes:
[0019] During the i-th round of data point selection, the data points selected from the current candidate data pool. The following formula is used to calculate:
[0020]
[0021] Where 1≤i≤c, and c is the total number of rounds in which data batches are selected; This indicates the b-th candidate in the current candidate data pool. i Data points, This represents the l-th data point in the current candidate data pool. This represents the data batch generated after i rounds of data point selection;
[0022] LB-BatchIG is a function for a set of data points, representing the reduction in the prediction uncertainty of the intelligent model for test samples after the set is added to the training dataset of the intelligent model;
[0023] This refers to The reduction in the prediction uncertainty of the intelligent model for test samples after the set of data points is added to the training dataset of the intelligent model is the amount of time that the model achieves this reduction. Indicates will Add the data batches generated in the previous round This then forms the data point set for the current round.
[0024] In one specific embodiment of the present invention, it further includes:
[0025] For any set of data points The calculation process is as follows:
[0026] First, a set of parameters {θ} is randomly sampled from the posterior distribution of the parameters of the intelligent model. l} 1≤l≤m θ l Let m represent the l-th parameter and m represent the number of sampling parameters;
[0027] Then calculate the two matrices. and These are the joint probability matrix and the marginal probability matrix, respectively, where K represents the number of categories the intelligent model needs to distinguish. The elements of these two matrices are calculated as follows:
[0028]
[0029] Where u and v represent the row and column numbers of the elements in the matrix, respectively; p(y=u|x,θ) l ) indicates that the model parameters are θ l Under the given conditions, the probability that the predicted result of x is y = u; x represents a sample generated from any data point in the candidate data pool. * Indicates a sample in the test environment; o u,v The prediction result of x is y = u and x * The probability of the predicted result y = v; qu,v This represents the probability of the predicted outcome y = u for x, and x. * The product of the probabilities of the predicted result y = v;
[0030] Calculate sample x against x * Information gain generated
[0031]
[0032] By testing the sample distribution p t (x * Upsampling x * Get a set The estimation results of the information gain metric criterion are as follows: s represents the number of test samples.
[0033] A second aspect of the present invention provides a model data hardening and optimization device based on information gain optimization, comprising:
[0034] The candidate data pool construction module is used to form a candidate data pool from data points that can be used as input data for the intelligent model, and to set the total number of rounds to select data points from the candidate data pool.
[0035] The data batch generation module is used to search for the data point that maximizes the improvement of the information gain metric estimation result from the current candidate data pool when each data point selection round arrives, and add it to the data batch saved in the previous round to obtain the current data batch; when the total number of data points in the current data batch is equal to the total number of rounds, the final data batch generation is completed.
[0036] The data hardening module is used to label each data point in the final data batch with corresponding labels to form new training samples, and then add all the new training samples to the original training dataset of the intelligent model to complete the data hardening of the training dataset.
[0037] In one specific embodiment of the present invention, it further includes:
[0038] During the i-th round of data point selection, the data points selected from the current candidate data pool. The following formula is used to calculate:
[0039]
[0040] Where 1≤i≤c, and c is the total number of rounds in which data batches are selected; This indicates the b-th candidate in the current candidate data pool. i Data points, This represents the l-th data point in the current candidate data pool. This represents the data batch generated after i rounds of data point selection;
[0041] LB-BatchIG is a function for a set of data points, representing the reduction in the prediction uncertainty of the intelligent model for test samples after the set is added to the training dataset of the intelligent model;
[0042] This refers to The reduction in the prediction uncertainty of the intelligent model for test samples after the set of data points is added to the training dataset of the intelligent model is the amount of time that the model achieves this reduction. Indicates will Add the data batches generated in the previous round This then forms the data point set for the current round.
[0043] In one specific embodiment of the present invention, it further includes:
[0044] For any set of data points The calculation process is as follows:
[0045] First, a set of parameters {θ} is randomly sampled from the posterior distribution of the parameters of the intelligent model. l} 1≤l≤m θ l Let m represent the l-th parameter and m represent the number of sampling parameters;
[0046] Then calculate the two matrices. and These are the joint probability matrix and the marginal probability matrix, respectively, where K represents the number of categories the intelligent model needs to distinguish. The elements of these two matrices are calculated as follows:
[0047]
[0048] Where u and v represent the row and column numbers of the elements in the matrix, respectively; p(y=u|x,θ) l ) indicates that the model parameters are θ l Under the given conditions, the probability that the predicted result of x is y = u; x represents a sample generated from any data point in the candidate data pool. * Indicates a sample in the test environment; o u,v The prediction result of x is y = u and x * The probability of the predicted result y = v; q u,v This represents the probability of the predicted outcome y = u for x, and x. * The product of the probabilities of the predicted result y = v;
[0049] Calculate sample x against x * Information gain generated
[0050]
[0051] By testing the sample distribution p t (x * Upsampling x * Get a set The estimation results of the information gain metric criterion are as follows: s represents the number of test samples.
[0052] A third aspect of the present invention provides an electronic device comprising:
[0053] At least one processor; and a memory communicatively connected to said at least one processor;
[0054] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to execute the above-described model data hardening optimization method based on information gain optimization.
[0055] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute the above-described model data hardening optimization method based on information gain optimization.
[0056] Features and beneficial effects of the present invention:
[0057] Existing techniques for hardening perception models from a model optimization perspective suffer from limitations such as limited training data and unstable optimization processes, resulting in less than ideal model hardening effects in practical applications. This invention addresses this by expanding training data to enhance model performance in testing environments without compromising the optimization method. However, existing data expansion methods focus on the properties of the data samples themselves and their relationship with model parameters, neglecting the impact of data samples on model performance under autonomous driving perception test samples. Furthermore, they ignore the distribution differences between the candidate sample data pool and the perception test sample distribution, preventing the expanded data samples from specifically hardening performance in the test environment. To address these shortcomings, this invention utilizes information entropy theory to establish an information gain metric that characterizes the degree to which various data batches in the candidate data pool improve the perception model's performance in the test environment. An efficient greedy method is designed to construct data batches, thereby completing the performance hardening of the target test environment.
[0058] This invention introduces the mutual information of prediction results between samples as the basis for calculating information gain in models with uncertainty. The expected maximum mutual information value of each sample in a data batch within the candidate data pool relative to samples in the test environment is used as the information gain metric. Based on the mathematical properties of this information gain metric, an efficient approximate solution method based on a greedy algorithm is established, achieving optimization of the information gain and diversity of the constructed autonomous driving perception sample batches.
[0059] This invention designs a candidate data batch pair model information gain metric criterion on the target test sample, which solves the problem that the original criteria ignore the inconsistency between the distribution of the target test sample and the distribution of the candidate pool data.
[0060] This invention introduces the goal of data sample diversity into the measurement criteria of data batches to solve the problem of redundant information.
[0061] Since the number of possible combinations of data batches is exponentially related to the batch size, this invention designs a data batch construction scheme with polynomial complexity to obtain the optimal data batch solution within a relatively short time overhead. This invention designs an information gain metric for the intelligent model on the target test samples based on data batches, while simultaneously considering both the reinforcement of prediction performance on the target test samples and the distribution mismatch of the data candidate pool.
[0062] This invention implicitly incorporates the objective of sample diversity in batches into the designed information gain metric, thereby significantly improving the performance of the model trained on the target test samples after obtaining data batches for model hardening based on this metric and adding them to the training dataset.
[0063] This invention designs an efficient construction scheme for hardened data batches. Based on the reasoning and analysis of the information gain measurement criterion, a multi-round iterative greedy construction scheme is designed to avoid the huge running overhead of exhaustively enumerating all potential batches.
[0064] This invention relates to the training data collection process for intelligent models of autonomous driving perception systems when optimizing performance under specific test environments (such as extreme weather conditions like rain, snow, and fog, and strong / weak lighting conditions). With the same amount of data collected, the trained model can achieve superior performance compared to other techniques in the test environment. Attached Figure Description
[0065] Figure 1 This is an overall flowchart of a model data hardening optimization method based on information gain optimization according to an embodiment of the present invention.
[0066] Figure 2This is a comparison chart showing the performance enhancement of the method described in this embodiment and other methods after constructing a visual perception task using the VLCS dataset in a specific embodiment of the present invention.
[0067] Figure 3 This is a comparison chart showing the performance enhancement of the method described in this embodiment and other methods after constructing a visual perception task using the PACS dataset in a specific embodiment of the present invention.
[0068] Figure 4 This is a comparison chart showing the performance enhancement effects of the method described in this embodiment and other methods on the model after constructing a visual perception task using the NICO++ dataset in a specific embodiment of the present invention. Detailed Implementation
[0069] This invention proposes a model data hardening optimization method and apparatus based on information gain optimization, which will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0070] A first aspect of this invention proposes a model data hardening optimization method based on information gain optimization, comprising:
[0071] The data points that can be used as input data for the intelligent model are formed into a candidate data pool, and the total number of rounds to select data points from the candidate data pool is set.
[0072] When each data point selection round arrives, the data point that maximizes the improvement of the information gain metric estimation result is searched from the current candidate data pool and added to the data batch saved in the previous round to obtain the current data batch; when the total number of data points in the current data batch is equal to the total number of rounds, the final data batch is generated.
[0073] Each data point in the final data batch is labeled with a corresponding tag to form a new training sample. Then, all the new training samples are added to the original training dataset of the intelligent model to complete the data hardening of the training dataset.
[0074] In a specific embodiment of the present invention, the model data reinforcement optimization method based on information gain optimization selects a batch of data from a candidate data pool to expand the training dataset of the intelligent model, so that the intelligent model retrained using the expanded training dataset achieves the greatest performance improvement on the test data. The overall process of this method is as follows: Figure 1 As shown, it includes the following steps:
[0075] 1) Initialize the current data batch as
[0076] 2) Set the total number of rounds for selecting data batches to c. Establish a candidate data pool. In this embodiment, the data points in the candidate data pool can be used as input data for the intelligent model. The data points in the candidate data pool can be obtained from the Internet or collected by sensors in a real driving environment.
[0077] In this embodiment, c is usually set between 10 and 100. The method described in this embodiment completes the data reinforcement process after going through c rounds of data point selection.
[0078] 3) Let i = 1.
[0079] 4) When selecting data points in the i-th round, find the data points from the current candidate data pool that maximize the improvement of the information gain metric estimation result and add them to the data batch generated in the (i-1)-th round. In this way, the data batch of the i-th round is generated. This indicates the data batch generated after i rounds of data point selection.
[0080] In the i-th round (1≤i≤c), the data points selected from the current candidate data pool. The following formula is used to calculate:
[0081]
[0082] in, This indicates the b-th candidate in the current candidate data pool. i Data points, This represents the l-th data point in the current candidate data pool. LB-BatchIG is a function for a set of data points, representing the reduction in the prediction uncertainty of the intelligent model for test samples after adding this set to the training dataset of the intelligent model. This refers to The reduction in the prediction uncertainty of the intelligent model for test samples after the set of data points is added to the training dataset of the intelligent model is the amount of time that the model achieves this reduction. Indicates will Add the data batches generated in the previous round This then forms the data point set for the current round.
[0083] Furthermore, for any set of data points The calculation process is as follows:
[0084] First, a series of parameters {θ} are randomly sampled from the posterior distribution of the parameters of the intelligent model (in this embodiment, the intelligent model is required to have the ability to output uncertainty, such as a Bayesian neural network or a neural network that uses Monte Carlo neuron dropout) (e.g., obtained using randomly dropped neurons). l} 1≤l≤m (θ lLet m represent the l-th parameter and m represent the number of sampling parameters (usually m is 50 or 100). Then, calculate the two matrices. and These are the joint probability matrix and the marginal probability matrix, respectively. K represents the number of categories the intelligent model needs to distinguish. For example, in a traffic light recognition task, there are three types of traffic lights: red, yellow, and green, so K = 3. The elements in these two matrices are calculated as follows:
[0085]
[0086] Where u and v represent the row and column numbers of the elements in the matrix, respectively. p(y=u|x,θ) l ) indicates that the model parameters are θ l Under the given conditions, the probability that x will predict y = u. x represents a sample generated from any data point in the candidate data pool. * Indicates a sample in the test environment, o u,v The prediction result of x is y = u and x * The probability of the predicted outcome y = v. u,v This represents the probability of the predicted outcome y = u for x, and x. * The product of the probabilities of the predicted result y = v.
[0087] Based on the above estimation results, the sample x can be calculated for x. * Information gain generated
[0088]
[0089] By testing the sample distribution p t (x * Upsampling x * Get a set The estimation results of the information gain metric criterion are as follows: s is the number of test samples, which is usually between 50 and 100.
[0090] 5) Remove from the current candidate data pool.
[0091] 6) Judgment:
[0092] If i < c, then let i = i + 1, and then return to step 4);
[0093] Otherwise, the final data batch Once generated, proceed to step 7).
[0094] 7) To Each data point is labeled with a corresponding tag to form a new training sample. Then, these c new training samples are added to the original training dataset of the intelligent model to complete the data reinforcement of the training dataset.
[0095] Then, by retraining the intelligent model using the updated training dataset, the model's performance can be effectively improved.
[0096] It should be noted that the method described in this embodiment can be repeated multiple times in practical applications.
[0097] In one specific embodiment of the present invention, a perception model is constructed based on ResNet, a well-known network architecture in the field of computer vision, and then the training dataset of this perception model is strengthened using the method described in this embodiment. This perception model can be used in the field of autonomous driving to classify objects that may appear on the road, helping the autonomous driving system perceive its surrounding environment.
[0098] In this embodiment, the perception model is applied to the VLCS dataset to classify objects in RGB images captured by the camera, including traffic elements such as cars and people, common animals such as dogs and birds, and obstacles such as chairs. Therefore, the joint probability matrix and the marginal probability matrix have K = 5 rows / columns. The dataset includes four data domains: LABELME, CALTECH, SUN, and PASCAL. Initial training and test data with different data distributions are created by adjusting the mixing ratio of the data domains.
[0099] Specifically, in this embodiment, the initial training data includes 400 LABELME images, 400 CALTECH images, and 400 SUN images, and the test data includes 1000 PASCAL images. The initial candidate data pool covers all data domains, including 1000 LABELME images, 400 CALTECH images, 1000 PASCAL images, and 1000 SUN images.
[0100] In this embodiment, by introducing a dropout method to randomly set the values of neurons to 0 in the neurons before the last fully connected layer of the ResNet, the posterior distribution of the model parameters is sampled.
[0101] The method described in this embodiment is compared with the Expected Prediction Information Gain Optimization (EPIG) technique, the uniform selection technique, and the Bayesian Active Learning technique based on diversified gradient representations (BADGE). Specifically, in this embodiment, the total number of rounds of data selection is set to c=50, that is, each method selects 50 data points from the candidate data pool, labels them, adds them to the original training dataset of the perceptual model, and then retrains the model.
[0102] Figure 2 This is a comparison chart showing the performance enhancement of the method described in this embodiment and other methods after constructing a visual perception task using the VLCS dataset in a specific embodiment of the present invention. Figure 2 The horizontal axis represents the number of samples selected from the candidate data pool, and the vertical axis represents the model's accuracy. EPIG represents the expected prediction information gain optimization technique. uniform represents the uniform selection technique. BADGE represents Bayesian active learning based on diverse gradient representations. LB-BatchIG represents the method described in this embodiment. As the number of samples selected from the candidate data pool and added to the training dataset increases, the prediction accuracy of the retrained model on the test data gradually improves. Compared to other methods, the method described in this embodiment brings the greatest improvement in accuracy.
[0103] Similarly, the method described in this embodiment is compared with the comparison method on the PACS and NICO++ datasets. The perceptual model is also built using ResNet, and the sampling method for the posterior distribution of the model parameters is the same, with c=50. The initial training data, test data, and candidate data pool consist of:
[0104] In the example of PACS data application
[0105] Initial training dataset: 400 photos + 400 art images + 400 sketches
[0106] Test dataset: 1000 cartoons
[0107] Candidate data pool: 1000 photos + 1000 art images + 1000 cartoon images + 1000 sketch images
[0108] NICO++ data application example:
[0109] Initial training dataset: 400 autumn images + 400 rock images + 400 dim images + 400 grass images
[0110] Test dataset: 1000 outdoor photos + 1000 water photos
[0111] Candidate data pool: 1000 autumn photos + 1000 rock photos + 1000 dim photos + 1000 grass photos + 1000 outdoor photos + 1000 water photos
[0112] Figure 3 This is a comparison chart showing the performance enhancement of the method described in this embodiment and other methods after constructing a visual perception task using the PACS dataset in a specific embodiment of the present invention. Figure 3 In the graph, the horizontal axis represents the number of samples selected from the candidate data pool, and the vertical axis represents the model's accuracy. EPIG represents the expected prediction information gain optimization technique. uniform represents the uniform selection technique. BADGE represents Bayesian active learning based on diverse gradient representations. LB-BatchIG represents the method described in this embodiment. Figure 2 The results conveyed are similar: given the same sample size, the method described in this embodiment brings the greatest improvement in model accuracy.
[0113] Figure 4 This is a comparison chart showing the performance enhancement effects of the method described in this embodiment and other methods on the model after constructing a visual perception task using the NICO++ dataset in a specific embodiment of the present invention. Figure 4 In the graph, the horizontal axis represents the number of samples selected from the candidate data pool, and the vertical axis represents the model's accuracy. EPIG represents the expected prediction information gain optimization technique. uniform represents the uniform selection technique. BADGE represents Bayesian active learning based on diverse gradient representations. LB-BatchIG represents the method described in this embodiment. Figure 2 The results conveyed are similar: given the same sample size, the method described in this embodiment brings the greatest improvement in model accuracy.
[0114] To implement the above embodiments, a second aspect of the present invention proposes a model data hardening and optimization device based on information gain optimization, comprising:
[0115] The candidate data pool construction module is used to form a candidate data pool from data points that can be used as input data for the intelligent model, and to set the total number of rounds to select data points from the candidate data pool.
[0116] The data batch generation module is used to search for the data point that maximizes the improvement of the information gain metric estimation result from the current candidate data pool when each data point selection round arrives, and add it to the data batch saved in the previous round to obtain the current data batch; when the total number of data points in the current data batch is equal to the total number of rounds, the final data batch generation is completed.
[0117] The data hardening module is used to label each data point in the final data batch with corresponding labels to form new training samples, and then add all the new training samples to the original training dataset of the intelligent model to complete the data hardening of the training dataset.
[0118] In one specific embodiment of the present invention, it further includes:
[0119] During the i-th round of data point selection, the data points selected from the current candidate data pool. The following formula is used to calculate:
[0120]
[0121] Where 1≤i≤c, and c is the total number of rounds in which data batches are selected; This indicates the b-th candidate in the current candidate data pool. i Data points, This represents the l-th data point in the current candidate data pool. This represents the data batch generated after i rounds of data point selection;
[0122] LB-BatchIG is a function for a set of data points, representing the reduction in the prediction uncertainty of the intelligent model for test samples after the set is added to the training dataset of the intelligent model;
[0123] This refers to The reduction in the prediction uncertainty of the intelligent model for test samples after the set of data points is added to the training dataset of the intelligent model is the amount of time that the model achieves this reduction. Indicates will Add the data batches generated in the previous round This then forms the data point set for the current round.
[0124] In one specific embodiment of the present invention, it further includes:
[0125] For any set of data points The calculation process is as follows:
[0126] First, a set of parameters {θ} is randomly sampled from the posterior distribution of the parameters of the intelligent model. l} 1≤l≤m θ l Let m represent the l-th parameter and m represent the number of sampling parameters;
[0127] Then calculate the two matrices. and These are the joint probability matrix and the marginal probability matrix, respectively, where K represents the number of categories the intelligent model needs to distinguish. The elements of these two matrices are calculated as follows:
[0128]
[0129] Where u and v represent the row and column numbers of the elements in the matrix, respectively; p(y=u|x,θ) l ) indicates that the model parameters are θ l Under the given conditions, the probability that the predicted result of x is y = u; x represents a sample generated from any data point in the candidate data pool. * Indicates a sample in the test environment; ou,v The prediction result of x is y = u and x * The probability of the predicted result y = v; q u,v This represents the probability of the predicted outcome y = u for x, relative to x. * The product of the probabilities of the predicted result y = v;
[0130] Calculate sample x against x * Information gain generated
[0131]
[0132] By testing the sample distribution p t (x * Upsampling x * Get a set The estimation results of the information gain metric criterion are as follows: s represents the number of test samples.
[0133] This allows for the greatest performance improvement of the intelligent model under the same sample selection constraints, providing important technical support for the application of autonomous driving perception models in complex and unseen test scenarios.
[0134] To implement the above embodiments, a third aspect of the present invention provides an electronic device, comprising:
[0135] At least one processor; and a memory communicatively connected to said at least one processor;
[0136] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to execute the above-described model data hardening optimization method based on information gain optimization.
[0137] To implement the above embodiments, a fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute the above-described model data hardening optimization method based on information gain optimization.
[0138] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.
[0139] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform a model data hardening optimization method based on information gain optimization according to the above embodiments.
[0140] Computer program code for performing the operations of this disclosure can be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, and C++, and conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0141] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0142] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0143] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.
[0144] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0145] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0146] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0147] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0148] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.
Claims
1. A model data hardening optimization method based on information gain optimization, characterized in that, include: The data points that can be used as input data for the intelligent model are formed into a candidate data pool, and the total number of rounds to select data points from the candidate data pool is set. When each data point selection round arrives, the data point that maximizes the improvement of the information gain metric estimation result is searched from the current candidate data pool and added to the data batch saved in the previous round, thus obtaining the current data batch; When the total number of data points in the current data batch equals the total number of rounds, the final data batch is generated. Each data point in the final data batch is labeled with a corresponding tag to form a new training sample. Then, all the new training samples are added to the original training dataset of the intelligent model to complete the data hardening of the training dataset.
2. The method according to claim 1, characterized in that, Also includes: During the i-th round of data point selection, the data points selected from the current candidate data pool. The following formula is used to calculate: Where 1≤i≤c, and c is the total number of rounds in which data batches are selected; This indicates the b-th candidate in the current candidate data pool. i Data points, This represents the l-th data point in the current candidate data pool. This represents the data batch generated after i rounds of data point selection; LB-BatchIG is a function for a set of data points, representing the reduction in the prediction uncertainty of the intelligent model for test samples after the set is added to the training dataset of the intelligent model; This refers to The reduction in the prediction uncertainty of the intelligent model for test samples after the set of data points is added to the training dataset of the intelligent model is the amount of time that the model achieves this reduction. Indicates will Add the data batches generated in the previous round This then forms the data point set for the current round.
3. The method according to claim 2, characterized in that, Also includes: For any set of data points The calculation process is as follows: First, a set of parameters {θ} is randomly sampled from the posterior distribution of the parameters of the intelligent model. l } 1≤l≤m θ l Let m represent the l-th parameter and m represent the number of sampling parameters; Then calculate the two matrices. and These are the joint probability matrix and the marginal probability matrix, respectively, where K represents the number of categories the intelligent model needs to distinguish. The elements of these two matrices are calculated as follows: Where u and v represent the row and column numbers of the elements in the matrix, respectively; p(y=u|x,θ) l ) indicates that the model parameters are θ l Under the given conditions, the probability that the predicted result of x is y = u; x represents a sample generated from any data point in the candidate data pool. * Indicates a sample in the test environment; o u,v The prediction result of x is y = u and x * The probability of the predicted result y = v; q u,v This represents the probability of the predicted outcome y = u for x, relative to x. * The product of the probabilities of the predicted result y = v; Calculate sample x against x * Information gain generated By testing the sample distribution p t (x * Upsampling x * Get a set The estimation results of the information gain metric criterion are as follows: s represents the number of test samples.
4. A model data reinforcement and optimization device based on information gain optimization, characterized in that, include: The candidate data pool construction module is used to form a candidate data pool from data points that can be used as input data for the intelligent model, and to set the total number of rounds to select data points from the candidate data pool. The data batch generation module is used to search for the data point that maximizes the improvement of the information gain metric estimation result from the current candidate data pool when the data point selection round arrives, and add it to the data batch saved in the previous round, thereby obtaining the current data batch. When the total number of data points in the current data batch equals the total number of rounds, the final data batch is generated. The data hardening module is used to label each data point in the final data batch with corresponding labels to form new training samples, and then add all the new training samples to the original training dataset of the intelligent model to complete the data hardening of the training dataset.
5. The apparatus according to claim 4, characterized in that, Also includes: During the i-th round of data point selection, the data points selected from the current candidate data pool. The following formula is used to calculate: Where 1≤i≤c, and c is the total number of rounds in which data batches are selected; This indicates the b-th candidate in the current candidate data pool. i Data points, This represents the l-th data point in the current candidate data pool. This represents the data batch generated after i rounds of data point selection; LB-BatchIG is a function for a set of data points, representing the reduction in the prediction uncertainty of the intelligent model for test samples after the set is added to the training dataset of the intelligent model; This refers to The reduction in the prediction uncertainty of the intelligent model for test samples after the set of data points is added to the training dataset of the intelligent model is the amount of time that the model achieves this reduction. Indicates will Add the data batches generated in the previous round This then forms the data point set for the current round.
6. The apparatus according to claim 5, characterized in that, Also includes: For any set of data points The calculation process is as follows: First, a set of parameters {θ} is randomly sampled from the posterior distribution of the parameters of the intelligent model. l } 1≤l≤m θ l Let m represent the l-th parameter and m represent the number of sampling parameters; Then calculate the two matrices. and These are the joint probability matrix and the marginal probability matrix, respectively, where K represents the number of categories the intelligent model needs to distinguish. The elements of these two matrices are calculated as follows: Where u and v represent the row and column numbers of the elements in the matrix, respectively; p(y=u|x,θ) l ) indicates that the model parameters are θ l Under the given conditions, the probability that the predicted result of x is y = u; x represents a sample generated from any data point in the candidate data pool. * Indicates a sample in the test environment; o u,v The prediction result of x is y = u and x * The probability of the predicted result y = v; q u,v This represents the probability of the predicted outcome y = u for x, relative to x. * The product of the probabilities of the predicted result y = v; Calculate sample x against x * Information gain generated By testing the sample distribution p t (x * Upsampling x * Get a set The estimation results of the information gain metric criterion are as follows: s represents the number of test samples.
7. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in any one of claims 1-3.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1-3.