Artificial intelligence model training method
By performing multiple rounds of augmentation and noise combination processing on the large model training dataset, combined with automatic augmentation strategies and reinforcement learning, the problems of generalization ability and accurate output in large model training are solved, achieving efficient training and accurate output of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING SECURITY UNION IT CO LTD
- Filing Date
- 2026-03-12
- Publication Date
- 2026-06-19
AI Technical Summary
How can we achieve accurate output through data stitching technology while ensuring the generalization ability of large models and avoiding problems such as model overfitting and decreased training efficiency?
By splicing and combining datasets obtained by processing the same data source in different ways, and combining automatic augmentation strategies and noise injection, reinforcement learning and automatic parameter tuning are used for training. The order and number of times the dataset is used in model training are adjusted, and multiple rounds of data augmentation and noise combination are introduced to improve the model's generalization ability.
It improves the generalization ability and output accuracy of large models, while avoiding overfitting and reduced training efficiency, and improves test accuracy by about 10%.
Smart Images
Figure CN122242626A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large model training technology, specifically to an artificial intelligence model training method. Background Technology
[0002] In machine learning and deep learning, model overfitting is a long-standing challenge. Overfitting leads to excellent performance on the training set but poor generalization ability, resulting in poor performance on unseen data. Data augmentation techniques are used to obtain new training samples to expand the dataset, thereby improving the model's generalization ability. However, excessive data augmentation may cause the model to learn incorrect features, thus affecting its performance. Therefore, how to utilize existing datasets and data augmentation techniques to achieve relatively accurate output while ensuring good generalization ability in large trained models has become a key consideration in training large models. Data concatenation techniques have been developed to expand datasets while maintaining the proportion of real data in the training, validation, and test sets. However, if the proportion of concatenated data in the training set is too high, it may affect model fairness. Therefore, how to use concatenated data and other similar data to train large models requires new research directions. Summary of the Invention
[0003] Therefore, the technical problem to be solved by the present invention is to provide an artificial intelligence model training method, which uses datasets obtained from the same data source but with different processing to train a large model and then tests and evaluates the trained large model. Then, based on the test and evaluation results, the order of use and the number of times different datasets participate in the training of the large model are set, and the large model is trained. This method ensures that the trained large model has good generalization ability and also has a relatively accurate output capability.
[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0005] Artificial intelligence model training methods include the following steps:
[0006] S10) Obtain the large model training dataset and divide the large model dataset into the first training set, the first validation set and the test set, and denote the first training set, the first validation set and the test set as the first dataset group;
[0007] S20) Perform data augmentation on the first training set and the first validation set obtained in step S10 respectively to obtain the second training set and the second validation set, and denote the second training set, the second validation set and the test set as the second dataset group;
[0008] S30) The first training set and the second training set are concatenated to obtain the third training set, the first validation set and the second validation set are concatenated to obtain the third validation set, and the third training set, the third validation set and the test set are denoted as the third dataset group;
[0009] S40) Combine the first training set and the second training set into the fourth training set, combine the first validation set and the second validation set into the fourth validation set, and denote the fourth training set, the fourth validation set and the test set as the fourth dataset group.
[0010] S50) The same large model is trained, validated, and tested sequentially using the first, second, third, and fourth dataset groups, respectively. The test results are scored and denoted as G1, G2, G3, and G4. The percentage of each dataset group in the total number of training iterations of the large model is then calculated using the following formula:
[0011] ;
[0012] In the formula, G i The test score is the result of training, validating, and testing the large model on the i-th dataset group; i can be any one of 1, 2, 3, and 4.
[0013] S60) Set the number of training sessions for the large model, then calculate the number of times each dataset group is used for training the large model according to the proportion determined in step S50 and round it down. Then select the corresponding dataset group in ascending order of G1, G2, G3 and G4 to train the large model for the corresponding number of sessions.
[0014] In the above-mentioned artificial intelligence model training method, in step S20, an automatic augmentation strategy is used to augment the data when augmenting the first training set.
[0015] In the above-described artificial intelligence model training method, during step S20, when performing data augmentation on the first training set, the data augmentation is performed on the first training set through the following operations:
[0016] S21) Set the number of data augmentations, n;
[0017] S22) The percentage of data that underwent the k-th data augmentation is calculated using the following formula:
[0018] ;
[0019] In the formula, k is a natural number greater than or equal to 2 and less than or equal to n;
[0020] S23) Perform data augmentation on the first training set according to the calculation results of step S22.
[0021] In the above-mentioned artificial intelligence model training method, in step S22, when k is greater than 3, a correction coefficient γ is introduced for p. k Make corrections so that the corrected p k Less than or equal to 45%.
[0022] The aforementioned artificial intelligence model training methods also include:
[0023] Step S70) Additional training of the large model: Inject noise with a proportion of less than or equal to 30% into the fourth training set to obtain the fifth training set. Then, use the large model trained in step S60 with the fifth training set, the fourth validation set, and the test set for additional training.
[0024] In the above-mentioned artificial intelligence model training method, in step S70, the injected noise includes random noise, injected original noise, and enhanced noise. The injected original noise is part or all of the original noise that is from the same source as the first dataset group, and the enhanced noise is the noise obtained after data enhancement of the original noise.
[0025] In the above-mentioned artificial intelligence model training method, the ratio of random noise, injected original noise, and enhanced noise is (2-6):(1-3):(1-5); when the injected noise is more than twice the original noise, the enhanced noise that meets the requirements is generated by performing multiple data augmentations on the original noise.
[0026] In the above-mentioned artificial intelligence model training method, the large model training in step S60 and the additional training of the large model after training in step S60 in step 70 are both carried out using reinforcement learning.
[0027] In the above-mentioned artificial intelligence model training method, in step S70, the parameters and hyperparameters are adjusted automatically when the large model is trained.
[0028] In the above-mentioned artificial intelligence model training method, in step S60, the parameters and hyperparameters are adjusted automatically when the large model is trained.
[0029] The technical solution of the present invention achieves the following beneficial technical effects:
[0030] This invention provides a method for training large models by constructing datasets for training large models. It obtains datasets based on real data by processing the original real data in different ways, and then constructs new datasets based on the impact of these datasets on the training of large models. The large model is then trained using the new datasets. This not only avoids the drawbacks of overusing data augmentation techniques, but also eliminates problems such as model overfitting and decreased training efficiency caused by an excessive proportion of data generated from real data. Attached Figure Description
[0031] Figure 1 This is a flowchart of the artificial intelligence model training process in this invention. Detailed Implementation
[0032] The present invention will be further explained below with reference to examples.
[0033] In the process of training an artificial intelligence model, dataset construction and large model training can be separated into independent steps, or they can be combined into one. The key point is the degree of correlation between dataset construction and large model training, that is, whether the training of the large model will have a feedback effect on dataset construction.
[0034] In view of this, the present invention integrates dataset construction and large model training into a single large model training process.
[0035] like Figure 1 As shown, the artificial intelligence model training method includes the following steps:
[0036] S10) Obtain the large model training dataset and divide the large model dataset into the first training set, the first validation set and the test set, and denote the first training set, the first validation set and the test set as the first dataset group;
[0037] S20) Perform data augmentation on the first training set and the first validation set obtained in step S10 respectively to obtain the second training set and the second validation set, and denote the second training set, the second validation set and the test set as the second dataset group;
[0038] S30) The first training set and the second training set are concatenated to obtain the third training set, the first validation set and the second validation set are concatenated to obtain the third validation set, and the third training set, the third validation set and the test set are denoted as the third dataset group;
[0039] S40) Combine the first training set and the second training set into the fourth training set, combine the first validation set and the second validation set into the fourth validation set, and denote the fourth training set, the fourth validation set and the test set as the fourth dataset group.
[0040] S50) The same large model is trained, validated, and tested sequentially using the first, second, third, and fourth dataset groups, respectively. The test results are scored and denoted as G1, G2, G3, and G4. The percentage of each dataset group in the total number of training iterations of the large model is then calculated using the following formula:
[0041] ;
[0042] In the formula, G iThe test score is the result of training, validating, and testing the large model on the i-th dataset group; i can be any one of 1, 2, 3, and 4.
[0043] S60) Set the number of training sessions for the large model, then calculate the number of times each dataset group is used for training the large model according to the proportion determined in step S50 and round it down. Then select the corresponding dataset group in ascending order of G1, G2, G3 and G4 to train the large model for the corresponding number of sessions.
[0044] Step S70) Additional training of the large model: Inject noise with a proportion of less than or equal to 30% into the fourth training set to obtain the fifth training set. Then, use the large model trained in step S60 with the fifth training set, the fourth validation set, and the test set for additional training.
[0045] In particular, the large model training in step S60 and the additional training of the large model trained in step S60 in step S70 are both carried out using reinforcement learning. Furthermore, in step S60, the parameters and hyperparameters are adjusted automatically during model training, and in step S70, the parameters and hyperparameters are adjusted automatically during model training.
[0046] Since the data in the dataset used for training artificial intelligence models may be applied to training different tasks, the methods for data augmentation vary. To avoid the inconvenience of manual operation and reduce the impact of human error, this invention employs an automatic augmentation strategy in step S20 when augmenting the first training set. A technician selects the appropriate data augmentation operation, sets the augmentation strategy, and then augments the first training set.
[0047] During the training of artificial intelligence models, it is often encountered that the trained model has poor adaptability to some specific new samples or unknown data. This may be because some data samples used in the training of artificial intelligence models are few. In order to increase the number of these data samples and the number of data samples related to these data samples, this invention achieves data augmentation of the first training set through multiple rounds of data augmentation.
[0048] Specifically, in step S20, when performing data augmentation on the first training set, the data augmentation on the first training set is performed through the following operations:
[0049] S21) Set the number of data augmentations, n;
[0050] S22) The percentage of data that underwent the k-th data augmentation is calculated using the following formula:
[0051] ;
[0052] In the formula, k is a natural number greater than or equal to 2 and less than or equal to n;
[0053] S23) Perform data augmentation on the first training set according to the calculation results of step S22.
[0054] Furthermore, in step S22, when k is greater than 3, a correction coefficient γ is introduced for p. k Make corrections so that the corrected p k Less than or equal to 45%.
[0055] The method involves first performing overall augmentation on the first training set, and then performing data augmentation on the augmented first training set according to steps S22 and S23. The data augmentation method used in steps S22 and S23 is a data reduction strategy, meaning that the amount of data augmented in each round is always less than the amount of data augmented in the previous round. This not only increases data diversity but also reduces the negative impact of over-augmentation on the model. This is because the total amount of data samples obtained through multiple rounds of data augmentation is not dominant in the entire second training set, and to reduce the negative impact of over-augmentation on the model, the data augmentation of the first validation set is only performed once.
[0056] To improve the generalization ability of artificial intelligence models, noise is often added to the training set during training. Typically, this added noise is random noise. While this invention also involves adding noise, it selects a combination of noise that includes random noise. Specifically, in step S70, the injected noise includes random noise, injected original noise, and enhanced noise. The injected original noise is part or all of the original noise originating from the same source as the first dataset. The enhanced noise is the noise obtained after data augmentation of the original noise, and the ratio of random noise, injected original noise, and enhanced noise is (2-6):(1-3):(1-5). When the injected noise is more than twice the original noise, multiple data augmentations are performed on the original noise to generate enhanced noise that meets the requirements. In this embodiment, the preferred ratio of random noise, injected original noise, and enhanced noise is 3:2:4. The original noise consists of data that was removed to obtain the first training set, the first validation set, and the test set.
[0057] In this invention, a combination of noise, including random noise, is used as injected noise. The purpose is to help the model better learn the intrinsic features of the data, strengthen the correlation between correlated data, and thus improve the generalization ability of the artificial intelligence model. Compared to an artificial intelligence model trained only on a training set containing random noise, the artificial intelligence model trained on a training set containing the noise combination of this invention shows an approximately 10% improvement in test accuracy.
[0058] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of the claims of this patent application.
Claims
1. An artificial intelligence model training method, characterized in that, Includes the following steps: S10) Obtain the large model training dataset and divide the large model dataset into the first training set, the first validation set and the test set, and denote the first training set, the first validation set and the test set as the first dataset group; S20) Perform data augmentation on the first training set and the first validation set obtained in step S10 respectively to obtain the second training set and the second validation set, and denote the second training set, the second validation set and the test set as the second dataset group; S30) The first training set and the second training set are concatenated to obtain the third training set, the first validation set and the second validation set are concatenated to obtain the third validation set, and the third training set, the third validation set and the test set are denoted as the third dataset group; S40) Combine the first training set and the second training set into the fourth training set, combine the first validation set and the second validation set into the fourth validation set, and denote the fourth training set, the fourth validation set and the test set as the fourth dataset group. S50) The same large model is trained, validated, and tested sequentially using the first, second, third, and fourth dataset groups, respectively. The test results are scored and denoted as G1, G2, G3, and G4. The percentage of each dataset group in the total number of training iterations of the large model is then calculated using the following formula: ； In the formula, G i The test score is the result of training, validating, and testing the large model on the i-th dataset group; i can be any one of 1, 2, 3, and 4. S60) Set the number of training sessions for the large model, then calculate the number of times each dataset group is used for training the large model according to the proportion determined in step S50 and round it down. Then select the corresponding dataset group in ascending order of G1, G2, G3 and G4 to train the large model for the corresponding number of sessions.
2. The artificial intelligence model training method according to claim 1, characterized in that, In step S20, an automatic augmentation strategy is used to augment the data on the first training set.
3. The artificial intelligence model training method according to claim 2, characterized in that, In step S20, when performing data augmentation on the first training set, the data augmentation on the first training set is performed through the following operations: S21) Set the number of data augmentations, n; S22) The percentage of data that underwent the k-th data augmentation is calculated using the following formula: ； In the formula, k is a natural number greater than or equal to 2 and less than or equal to n; S23) Perform data augmentation on the first training set according to the calculation results of step S22.
4. The artificial intelligence model training method according to claim 3, characterized in that, In step S22, when k is greater than 3, a correction coefficient γ is introduced for p. k Make corrections so that the corrected p k Less than or equal to 45%.
5. The artificial intelligence model training method according to claim 1, characterized in that, Also includes: Step S70) Additional training of the large model: Inject noise with a proportion of less than or equal to 30% into the fourth training set to obtain the fifth training set. Then, use the large model trained in step S60 with the fifth training set, the fourth validation set, and the test set for additional training.
6. The artificial intelligence model training method according to claim 5, characterized in that, In step S70, the injected noise includes random noise, injected original noise, and enhanced noise. The injected original noise is part or all of the original noise that is from the same source as the first dataset group, and the enhanced noise is the noise obtained after the original noise has been data enhanced.
7. The artificial intelligence model training method according to claim 6, characterized in that, The ratio of random noise, injected original noise, and enhanced noise is (2-6):(1-3):(1-5); when the injected noise is more than twice the original noise, the enhanced noise that meets the requirements is generated by performing multiple data enhancements on the original noise.
8. The artificial intelligence model training method according to claim 7, characterized in that, The large model training in step S60 and the additional training of the large model trained in step S60 in step 70 are both carried out using reinforcement learning.
9. The artificial intelligence model training method according to claim 8, characterized in that, In step S70, the large model is trained using an automatic parameter tuning method to adjust parameters and hyperparameters.
10. The artificial intelligence model training method according to claim 8, characterized in that, In step S60, the large model is trained using an automatic parameter tuning method to adjust parameters and hyperparameters.