Active learning method and training system based on neural network regression

Abstract

The present invention relates to an active learning method and training system based on neural network regression, which uses labeled sample data for supervised training, selects high-quality unlabeled sample data for experiments based on the training results and the distribution of unlabeled sample data and labeled sample data, and adds them to the labeled data set for further training. The selection process fully considers the model information and sample distribution, and achieves the goal of using the smallest possible budget to maximize the accuracy of neural network model prediction.

Description

Technical Field

[0001] The present invention relates to a data processing technology, and in particular to an active learning method and a training system based on neural network regression. Background Art

[0002] With the rapid development of artificial intelligence in the computer field, the application of artificial intelligence in the field of materials science is becoming more and more extensive. Among them, material property prediction based on neural networks has achieved great success in the field of materials science.

[0003] In the field of materials science, the practicality of a certain material is judged based on its performance parameters. However, the performance parameters of materials are difficult to obtain directly and require complex and precise experiments to obtain. Using artificial intelligence technology to predict the performance parameters of materials can reduce or avoid a large number of costly experiments. When changing the composition or structure to obtain a new material, it is only necessary to measure the process parameters of the material and input them into the neural network to obtain a prediction of the performance parameters of the material and provide auxiliary judgment. However, the training of neural networks often requires the use of a large amount of high-quality data, otherwise the trained neural network may predict results that deviate greatly from the actual situation and cannot provide auxiliary judgment. Therefore, materials scientists need to experiment on a large number of materials to obtain performance parameters. However, the time cost and economic cost of experiments are generally high. How to maximize the prediction accuracy of the neural network model while minimizing the experimental cost is a difficult problem to solve.

[0004] Due to the influence of measurement methods and accuracy, the process parameter data of materials has a lot of noise, so randomly selecting materials for experiments cannot meet the above requirements. Active learning combines model information and data distribution to select as little high-quality unlabeled material data as possible, and after manual annotation, it is added to the training set to train the task model. Therefore, it is necessary to use active learning based on neural network regression for artificial intelligence material performance parameter prediction. Summary of the invention

[0005] In order to solve the problem of large data volume and high experimental cost, an active learning method and training system based on neural network regression was proposed. It can select high-quality samples from a large number of unlabeled samples for annotation, thereby expanding the labeled data set to train the regression model and improve the accuracy of regression model prediction.

[0006] The technical solution of the present invention is: an active learning method based on neural network regression, which specifically includes the following steps:

[0007] S100, establish task regression model:

[0008] There will be labeled data D lAs a training set, the task regression model is trained to fit or roughly fit the data quantitatively;

[0009] From the unlabeled dataset D l The trained task regression model is used to predict y1;

[0010] S200. Based on the trained task regression model, an auxiliary classification model is further established to perform outlier detection on unlabeled data:

[0011] Use the trained task regression model to predict the labeled data, train the auxiliary classification model based on the predicted results and the true labels, and then use the trained auxiliary classification model to classify the unlabeled dataset D u Classify and obtain the outlier probability of the input unlabeled data, and select the K unlabeled data with the lowest outlier probability to form the data set D k , and at the same time in y1 according to the subset D K Corresponding to the discrete point subset y 1K ;

[0012] S300, greedy sampling in the selected unlabeled dataset:

[0013] In K unlabeled datasets D k In the data geometry distance, select the labeled data D l The M unlabeled datasets D with the longest average distance from the sampled unlabeled data M , while in y 1K According to the subset D M Correspondingly, the greedy sampling subset y is also taken 1M ;

[0014] S400, greedy sampling and minimum change sampling of the data set:

[0015] M unlabeled datasets D M The labeled dataset D l The trained task regression model is used to predict y2, the sampling change |y 1M -y2|The smallest B unlabeled datasets D B ;

[0016] S500, the sampled B unlabeled data sets D B Conduct experiments to obtain the true labels of the unlabeled dataset and add them to the labeled dataset D l In the process, step S100 is returned and repeatedly executed to perform the next round of active learning cycle until the specified number of cycles is reached or all the experimental budgets are used up.

[0017] Furthermore, the training sample data used by the auxiliary classification model in S200 is labeled data, and the label y calculation formula of the auxiliary classification model output data is y=norm(|y pred -y*|)

[0018] Where norm(·) represents the normalization operation, y pred is the predicted output of the task regression model for the labeled data, y* is the original true label of the labeled data, and the absolute difference of the data is normalized to represent the probability of an outlier.

[0019] Furthermore, the auxiliary classification model has three inputs, namely, input original data features of the same dimension, primary features, and secondary features, wherein the primary features and secondary features are features extracted from different depths of the task regression model, respectively. Then, these three different levels of features are concatenated after passing through the hidden layer, and a binary classifier is connected after retaining the feature information to the maximum extent. The output data is the probability of an outlier.

[0020] Further, in S300, the D K Sampling is performed to obtain D M , sampling one sample at a time to ensure that adjacent feature spaces are not sampled multiple times, and the maximum guarantee D M diversity.

[0021] Furthermore, the specific method of iterative sampling is as follows: unlabeled data subset D M Initially, it is an empty set. Each iteration starts from D K Sampling a labeled dataset D l and the unlabeled data subset D M The farthest sample, iterate M times to get the unlabeled data subset D M , D M The data samples in are both diverse and not outliers.

[0022] Furthermore, the unlabeled data sampled in S400 predicts the data samples with the smallest change before and after the training task regression model. The obtained data samples are all after steps S200 and S300. The obtained data samples are the sparse part of the existing labeled data set, while maintaining the diversity of the original labeled data.

[0023] Furthermore, in S400, the data sample with the smallest prediction change before and after the task regression model training is selected in one cycle, and the step S100 of the next cycle will continue to train the task regression model, resulting in changes in the task regression model in this part of the feature space. Then, in the step S400 of the next cycle, data from other parts of the feature space will be sampled, thereby ensuring that the task regression model can accurately predict all data after multiple cycles.

[0024] An active learning training system based on neural network regression uses labeled data and unlabeled data. The task regression model and auxiliary classification model are combined to learn the features of labeled data. An active learning method with a discrete point removal strategy is adopted to select high-quality materials from a large amount of unlabeled data materials for experiments, thereby expanding the diversity of labeled data sets used to train task regression models and improving the accuracy of task regression model predictions in iterative training.

[0025] The beneficial effects of the present invention are as follows: the present invention is based on an active learning method and training system for neural network regression, uses labeled sample data for supervised training, selects high-quality unlabeled sample data for experiments based on the training results and the distribution of unlabeled sample data and labeled sample data, and adds the labeled data set to train again. The selection process fully considers the model information and sample distribution, and achieves the goal of using as little budget as possible to maximize the accuracy of neural network model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of an embodiment of an active learning method based on neural network regression of the present invention;

[0027] Figure 2 It is a schematic diagram of data flow in the method of the present invention;

[0028] Figure 3 The figure is a schematic diagram of the structure and relationship of the task model and outlier detection module in the method of the present invention. DETAILED DESCRIPTION

[0029] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0030] like Figure 1 The flowchart of the active learning method based on neural network regression is shown, which selects a batch of material experiments cyclically until all budgets are used up, including the following steps:

[0031] S100, establish task regression model: have labeled data D l As a training set, the task regression model is trained to fit or roughly fit, quantitatively describe the data, and enable the task regression model to more accurately predict the labeled data;

[0032] S200, based on the trained task regression model, further establish an auxiliary classification model to detect outliers on the unlabeled data: use the trained and fitted task regression model to predict the labeled data, train the auxiliary classification model based on the prediction results and the real labels, so that the auxiliary classification model can more accurately distinguish whether the labeled data is an outlier, and then use the trained auxiliary classification model to detect outliers on the unlabeled data set D u Classify and obtain the outlier probability of the input unlabeled data, and select the K unlabeled data with the lowest outlier probability to form the data set D k ;

[0033] S300, greedy sampling in the selected unlabeled dataset: in K unlabeled datasets D k In the data geometry distance, select the labeled data D l The M unlabeled datasets D with the longest average distance from the sampled unlabeled data M ;

[0034] S400, after greedy sampling, the dataset is sampled with minimum change: from the unlabeled dataset D l The trained task regression model predicts y1, which corresponds to D k , D M The prediction result y obtained 1M , M unlabeled datasets D M The labeled dataset D l The trained task regression model is used to predict y2, the sampling change |y 1M -y2|The smallest B unlabeled datasets D B ;

[0035] S500, the sampled B unlabeled data sets D B Conduct experiments to obtain the true labels of the unlabeled dataset and add them to the labeled dataset D l In the process, step S100 is returned and repeatedly executed to perform the next round of active learning cycle until the specified number of cycles is reached or all the experimental budgets are used up.

[0036] In this embodiment, in step S100, before training the task regression model, it is necessary to use the task regression model to predict the unlabeled data set D u The output of the prediction output is selected in accordance with the results of steps S200 and S300, and then used to determine the change size of the unlabeled data in step S400.

[0037] In this embodiment, in step S200, the sample data for training the auxiliary classification model is labeled data, and the label y calculation formula of the auxiliary classification model output data is:

[0038] y=norm(|y pred -y*|)

[0039] Where norm(·) represents the normalization operation, normalizing the absolute difference of the data to represent the probability of an outlier, y pred is the predicted output of the task regression model for the labeled data, and y* is the original true label of the labeled data.

[0040] In this embodiment, in step S300, a greedy sampling algorithm is used to select D K D M The above algorithm samples in an iterative form, sampling one sample each time to ensure that adjacent feature spaces are not sampled multiple times, thereby maximizing the possibility of guaranteeing D M diversity.

[0041] In this embodiment, an active learning model training system based on neural network regression includes a task regression model and an auxiliary classification model for detecting outliers, and an initial labeled data set is used as a training set for the model.

[0042] The following lists the preferred embodiments of the active learning model training system based on neural network regression to clearly illustrate the content of the present invention. It should be clear that the content of the present invention is not limited to the following embodiments, and other improvements through conventional technical means of ordinary technicians in this field are also within the scope of the concept of the present invention.

[0043] The embodiment of the present invention proposes an active learning model training method based on neural network regression, and specifically describes each step in detail.

[0044] S100, task regression model training: There is labeled data D in the training set l Train the task regression model to fit or roughly fit, so that the task model can predict the labeled data more accurately;

[0045] Specifically, before training the task regression model, we first need to train the unlabeled dataset D u A prediction is made and the output is y1, which is used for the minimum change sampling in step S400.

[0046] S200, outlier detection: Use the trained and fitted task model to predict the labeled data, and train the auxiliary classification model based on the predicted results and the real labels, so that the auxiliary classification model can more accurately distinguish whether the labeled data is an outlier. Then use the trained auxiliary model to detect the unlabeled data set D u Classify and get the probability that the unlabeled data is an outlier. Select the K unlabeled data with the lowest outlier probability to form the data set D k ;

[0047] Specifically, the auxiliary classification model, like the task regression model, must be trained in each active learning cycle. Figure 2 As shown in the figure, the training data of the auxiliary classification model for outlier detection is labeled data, and the label is the absolute difference between the normalized prediction and the true label. Then the trained auxiliary classification model is used to classify the unlabeled data set D u Perform classification and select the subset D with the smallest probability of outliers K , and at the same time in y1 according to the subset D K Corresponding to the discrete point subset y 1K .

[0048] Specifically, Figure 3 As shown in the figure, the auxiliary classification model has three inputs, namely the original data features, primary features, and secondary features of the same dimension. The primary features and secondary features are features extracted from different depths of the task regression model respectively. Then, these three different levels of features are concatenated after passing through the hidden layer. After retaining the feature information to the maximum extent, a binary classifier is connected, and the output data is the probability of an outlier.

[0049] S300, Greedy Sampling: In K unlabeled datasets D k In the above example, select M unlabeled data sets D that have the longest average distance from labeled data and selected unlabeled data in terms of data geometry distance. M ;

[0050] Specifically, as mentioned above, this step is an iterative process. M Initially, it is an empty set. Each iteration starts from D K Sampling a labeled dataset D l and the unlabeled data subset D M The farthest sample, iterate M times to get the unlabeled data subset D M . D M The data samples in y are both diverse and not outliers. 1K According to the subset D M Correspondingly, the greedy sampling subset y is also taken 1M .

[0051] S400, minimum change sampling: in M ​​unlabeled datasets D M On the labeled dataset D l The trained task model predicts y2, the sampling change |y 1M -y2|The smallest B unlabeled datasets D B ;

[0052] Specifically, y 1MIt has been calculated in step S100 and sampled in steps S200 and S300, so it is only necessary to M The fitted task regression model is used to predict y2, and then the minimum change sampling is performed. B Yes D M A subset of M The model change information is taken into account on the basis of

[0053] S500, the sampled B unlabeled data sets D B Conduct experiments to obtain the true labels of the unlabeled dataset and add them to the labeled dataset D l Repeat step S100 to perform the next round of active learning cycle until the specified number of cycles is reached or all experimental budgets are used up.

[0054] The present invention uses an active learning method that combines sample data diversity, task regression model information and discrete point removal strategy to select high-quality materials from a large amount of unlabeled data materials for experiments, expand the labeled data set used to train the task regression model, and greatly improve the accuracy of task regression model prediction while using as little budget as possible.

[0055] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An active learning method based on neural network regression, characterized in that: A batch of material experiments are selected cyclically until all budgets are used up. The method specifically includes the following steps: S100, establish task regression model: There will be labeled data D l As a training set, the task regression model is trained to fit or roughly fit the data quantitatively; From the unlabeled dataset D l The trained task regression model is used to predict y1; S200. Based on the trained task regression model, an auxiliary classification model is further established to perform outlier detection on unlabeled data: Use the trained task regression model to predict the labeled data, train the auxiliary classification model based on the predicted results and the true labels, and then use the trained auxiliary classification model to classify the unlabeled dataset D u Classify and obtain the outlier probability of the input unlabeled data, and select the K unlabeled data with the lowest outlier probability to form the data set D k , and at the same time in y1 according to the subset D K Corresponding to the discrete point subset y 1K ; S300, greedy sampling in the selected unlabeled dataset: In K unlabeled datasets D k In the data geometry distance, select the labeled data D l The M unlabeled datasets D with the longest average distance from the sampled unlabeled data M , while in y 1K According to the subset D M Correspondingly, the greedy sampling subset y is also taken 1M ; S400, greedy sampling and minimum change sampling of the data set: M unlabeled datasets D M The labeled dataset D l The trained task regression model is used to predict y2, the sampling change |y 1M -y2|The smallest B unlabeled datasets D B ; S500, the sampled B unlabeled data sets D B Conduct experiments to obtain the true labels of the unlabeled dataset and add them to the labeled dataset D l In the process, step S100 is returned and repeatedly executed to perform the next round of active learning cycle until the specified number of cycles is reached or all the experimental budgets are used up.

2. The active learning method based on neural network regression according to claim 1, characterized in that: The training sample data used by the auxiliary classification model in S200 is labeled data, and the label y calculation formula of the auxiliary classification model output data is y=norm(y pred -y*) Where norm(g) represents the normalization operation, y pred is the predicted output of the task regression model for the labeled data, y* is the original true label of the labeled data, and the absolute difference of the data is normalized to represent the probability of an outlier.

3. The active learning method based on neural network regression according to claim 2 is characterized in that: The auxiliary classification model has three inputs, namely, input original data features of the same dimension, primary features, and secondary features. The primary features and secondary features are features extracted from different depths of the task regression model, respectively. These three different levels of features are then concatenated after passing through the hidden layer, and a binary classifier is connected after retaining the feature information to the maximum extent. The output data is the probability of an outlier.

4. The active learning method based on neural network regression according to claim 1, characterized in that: In S300, the D K Sampling is performed to obtain D M , sampling one sample at a time to ensure that adjacent feature spaces are not sampled multiple times, and the maximum guarantee D M diversity.

5. The active learning method based on neural network regression according to claim 4 is characterized in that: The specific method of sampling in the iterative form: unlabeled data subset D M Initially, it is an empty set. Each iteration starts from D K Sampling a labeled dataset D l and the unlabeled data subset D M The farthest sample, iterate M times to get the unlabeled data subset D M , D M The data samples in are both diverse and not outliers.

6. The active learning method based on neural network regression according to any one of claims 1 to 5, characterized in that: The unlabeled data sampled in S400 has the smallest predicted change before and after the training task regression model. The obtained data samples are all after steps S200 and S300. The obtained data samples are the sparse part of the existing labeled data set, while maintaining the diversity of the original labeled data.

7. The active learning method based on neural network regression as claimed in claim 6, characterized in that: In S400, the data sample with the smallest prediction change before and after the task regression model training is selected in one cycle, and step S100 of the next cycle will continue to train the task regression model, resulting in changes in the task regression model in this part of the feature space. In the next cycle, step S400 will sample data from other parts of the feature space to ensure that the task regression model can accurately predict all data after multiple cycles.

8. An active learning training system based on neural network regression, characterized in that: Active learning training is performed using the active learning method based on neural network regression as described in any one of claims 1 to 7, using labeled data and unlabeled data, and combining the task regression model and the auxiliary classification model to learn the features of the labeled data. An active learning method with a discrete point removal strategy is adopted, and high-quality materials are selected from a large amount of unlabeled data materials for experiments, thereby expanding the diversity of labeled data sets used to train the task regression model, and improving the accuracy of the task regression model prediction in iterative training.

Images