TWI840640B - Semi-supervised learning system and semi-supervised learning method - Google Patents

Abstract

A semi-supervised learning method is provided. The method includes the following steps: obtaining source-domain data of one or more source domains and target-domain data of a target domain; training a feature-extraction model using the source-domain data and the target-domain data; calculating a domain-determination loss function, a task loss function, and a semi-supervised loss function respectively by a domain-determination model, a task model, and a semi-supervised learning mechanism; calculating a total loss function according to the domain-determination loss function, the task loss function, and the semi-supervised loss function, and updating weights of the feature-extraction model according to the total loss function; and finishing a training procedure of the an overall model in response to the overall model satisfying a model converge condition, wherein the overall model includes the feature-extraction model, the task model, and the domain-determination model.

Description

Semi-supervisory learning system and semi-supervisory learning method

The present invention relates to machine learning, and more particularly to a semi-supervised learning system and a semi-supervised learning method.

After collecting a large amount of labeled data in a specific application domain (hereinafter referred to as the source domain), the data-oriented analytical modeling technology can train the model to perform predictions close to or even better than humans in the same domain. However, when we want to reuse the model in another new domain (hereinafter referred to as the target domain) for the same prediction task, the data set in the new domain often has a distribution gap with the past training data, resulting in a significant reduction in the model's prediction performance. For example, if the same product is produced in multiple factories, the appearance defect recognition system established using the production data of Factory A, if directly applied to Factory B for recognition, may have differences in data distribution due to differences in shooting angles, lighting, camera models, etc., resulting in less than expected operating results.

In view of this, the present invention provides a semi-supervised learning system and a semi-supervised learning method to solve the above problems.

The present invention provides a semi-supervised learning system, comprising: a non-volatile memory for storing a semi-supervised learning application; and a processor for executing the semi-supervised learning application to execute the following steps: obtaining source domain data of one or more source domains and target domain data of a target domain; using the source domain data and the target domain data to train a feature extraction model; using a domain discrimination model, a task model and a semi-supervised learning mechanism to respectively calculate the feature extraction model; A domain discrimination loss function, a task loss function and a semi-supervisory loss function are included in the training process; a total loss function is calculated based on the domain discrimination loss function, the task loss function and the semi-supervisory loss function, and the weights of the feature extraction model, the task model and the domain discrimination model are updated based on the total loss function; and in response to the overall model satisfying the model convergence condition, the training process of the overall model is terminated, wherein the overall model includes the feature extraction model, the task model and the domain discrimination model.

In some embodiments, the feature extraction model is a ResNet50 model, the semi-supervised learning mechanism is an unsupervised data augmentation mechanism with consistency regularization, the task model is a first fully connected layer, and the domain discrimination model is a second fully connected layer plus a gradient reversal layer.

In some embodiments, the processor assigns a first hyper parameter, a second hyper parameter, and a third hyper parameter to the domain determination loss function, the task loss function, and the semi-supervisory loss function, respectively, to calculate the total loss function.

In some embodiments, the source domain data includes first labeled data and first unlabeled data, and the target domain data includes second labeled data and second unlabeled data. The processor uses the first labeled data, the first unlabeled data, the second labeled data, and the second unlabeled data to calculate the domain discriminant function, and updates the weights of the feature extraction model and the domain discriminant model according to the domain discriminant loss function. The processor uses the first labeled data and the second labeled data to calculate the task loss function, and updates the weight of the feature extraction model and the second weight of the task model according to the task loss function. The processor uses the first unlabeled data and the second unlabeled data to calculate the semi-supervised loss function, and updates the weights of the feature extraction model according to the semi-supervised loss function. In addition, the model convergence condition indicates that the improvement of the total loss function of the overall model under a fixed training batch (epoch) and a fixed training time is lower than a critical value.

The present invention further provides a semi-supervised learning method, comprising: obtaining source domain data of one or more source domains and target domain data of a target domain; using the source domain data and the target domain data to train a feature extraction model; using a domain discrimination model, a task model and a semi-supervised learning mechanism to respectively calculate the domain discrimination loss function, the task loss function and the semi-supervised loss function of the feature extraction model; and The task loss function and the semi-supervisory loss function are used to calculate a total loss function, and the first weight, the second weight and the third weight corresponding to the feature extraction model, the task model and the domain discrimination model are updated according to the total loss function; and in response to an overall model satisfying a model convergence condition, the training process of the overall model is terminated, wherein the overall model includes the feature extraction model, the task model and the domain discrimination model.

The following description is a preferred embodiment of the invention, and its purpose is to describe the basic spirit of the invention, but it is not intended to limit the invention. The actual content of the invention must refer to the scope of the following claims.

It must be understood that the words "comprise", "include" and the like used in this specification are used to indicate the existence of specific technical features, numerical values, method steps, operation processes, elements and/or components, but do not exclude the addition of more technical features, numerical values, method steps, operation processes, elements, components, or any combination thereof.

The terms "first", "second", "third", etc. used in the claims are used to modify the elements in the claims and are not used to indicate a priority order, a preceding relationship, or that one element precedes another element, or a temporal sequence in performing method steps. They are only used to distinguish elements with the same name.

FIG. 1 is a schematic diagram of a semi-supervised learning system according to an embodiment of the present invention.

The semi-supervised learning system 100 includes one or more processors 110, a memory unit 130, a storage device 140, and a transmission interface 150. The processing unit 110 may be, for example, a central processing unit (CPU), a general-purpose processor, etc., but the present invention is not limited thereto.

The memory unit 130 is a random access memory, such as a dynamic random access memory (DRAM) or a static random access memory (SRAM), but the present invention is not limited thereto. The storage device 140 is a non-volatile memory, such as a hard disk drive, a solid-state disk, a flash memory, or a read-only memory, but the present invention is not limited thereto.

For example, the storage device 140 can store a feature extraction model 141, a semi-supervised learning mechanism 142, a task model 143, and a domain discrimination model 144, and can be collectively referred to as a semi-supervised learning application (for example, the semi-supervised learning method of the present invention can be executed). The processor 110 reads the feature extraction model 141, the semi-supervised learning mechanism 142, the task model 143, and the domain discrimination model 144 into the memory unit 130 and executes them. In some embodiments, the storage device 140 may further store source domain data and target domain data obtained from different source domains and target domains, and the source domain data and target domain data include labeled and unlabeled data.

The transmission interface 150 may be, for example, a wired transmission interface or a wireless transmission interface, and the semi-supervised learning system 100 may be connected to one or more external devices 20 via the transmission interface 150 and receive source domain data or target domain data from the external devices.

In one embodiment, the semi-supervised learning system 100 combines the advantages of domain adaptation technology and semi-supervised learning technology. For example, domain adaptation technology can learn domain invariant universal features from data in the source domain and the target domain, and establish a discriminant model from the relationship between these universal features and source domain labels. Since the model makes judgments based on domain invariant features, it is less likely to be affected by changes in the distribution of target domain data, and thus can be used in the target domain for prediction. If the training data only contains a small amount of labeled data and a large amount of unlabeled data, the relevant technologies of supervised learning and unsupervised learning are usually integrated in the learning step of the model. The performance of an excellent semi-supervised learning model will be greatly improved compared to when only unsupervised learning is used, and it is close to the performance of a supervised learning model that uses only labeled data, which greatly saves the cost of labeling. Therefore, the semi-supervised learning system 100 of the present invention can better learn universal features that are domain-invariant and effective for the target domain through domain adaptation technology and the mechanism of semi-supervised learning, and can still achieve good model performance when the amount of source domain labeled data or the amount of target domain labeled data is small.

It should be noted that the present invention does not limit the model used by the semi-supervised learning system 100, and the above-mentioned model can be a method such as machine learning, statistical model and deep learning. In addition, the semi-supervised learning system 100 does not limit the type of training data used, which may include but not limited to structured data, signal data, image pictures, text data, etc. The training method of the above-mentioned model is not limited to stages, and different models can be trained separately in multiple stages or the entire end-to-end model can be used for training at the same time.

When both the source domain data and the target domain data are images, VGG, ResNet or Inception models can be used as feature extraction models 141. In a preferred embodiment, the semi-supervised learning system 100 uses ResNet50 as the feature extraction model 141, wherein the feature extraction model 141 can extract abstract features of source images (e.g., unlabeled images in the source domain data and the target domain data), which can be represented by feature vectors, for example, and can be used for subsequent defect detection processing.

The semi-supervised learning mechanism 142 may be, for example, an entropy-based regularization technique or a consistency regularization technique. In a preferred embodiment, the semi-supervised learning mechanism 142 is an unsupervised data augmentation mechanism using consistency regularization, which may utilize unlabeled samples to calculate information consistency to ensure that the task model is capable of defect discrimination on unknown samples.

The task model 143 may be, for example, a fully connected layer, which may map the feature vector generated by the feature extraction model 143 to generate a classification result.

The domain discrimination model 144 can be, for example, a fully connected layer plus a gradient reversal layer, or a generative adversarial network. In a preferred embodiment, the domain discrimination model 144 uses a fully connected layer plus a gradient reversal layer, which can ensure that the abstract features learned by the feature extraction model are domain invariant, so as to have a more robust judgment ability when facing data from different source domains.

FIG. 2 is a schematic diagram of the training process of the feature extraction model according to an embodiment of the present invention.

In one embodiment, the training information of the feature extraction model 141 includes source domain data 210~21N and target domain data 220, and each source domain data 210~21N and target domain data 220 respectively includes labeled data and unlabeled data. For example, the source domain data 210 includes labeled data 210A and unlabeled data 210B, the source domain data 211 includes labeled data 211A and unlabeled data 211B, and so on.

In one embodiment, the processor 110 uses the source domain data 210-21N and the unlabeled data 210B-21NB and 220B in the target domain data 220 to train the feature extraction model 141, for example, self-supervised learning or a generative adversarial network may be used, and the feature extraction model 141 may extract abstract descriptive features common to the source domain and the target domain.

The semi-supervised learning system 100 of the present invention can be used in a variety of fields, such as face recognition, object detection, image speech segmentation, text data (for example, can be used for dialogue robots and article abstracts), signal anomaly detection, component life prediction, etc., but the present invention is not limited to the application in the above fields. Therefore, according to the data type of different applications and the method of obtaining sensor data, the source domain data and the target domain data are, for example, data that have been pre-processed to be input into the feature extraction model 141 for training. Therefore, the domain discrimination model 144 can be established using the source domain data and the target domain data, which can ensure that the features extracted from different source domain data have a common expression space.

Next, the processor 110 uses the labeled data in the source domain data 210-21N and the target domain data 220 to train the task model 142 to generate a plurality of feature vectors 231. In addition, the semi-supervised learning mechanism 143 uses the unlabeled data in the source domain data and the target domain data to assist the task model training, and its purpose is to find an effective extraction method from the unlabeled data.

During the training process of the feature extraction model 141, the processor 110 continuously calculates the domain discrimination model 144, the task model 143 and the loss function of the semi-supervised learning mechanism 142 according to the source domain data and the target domain data, such as the domain discrimination loss function 234, the task loss function 233 and the semi-supervised loss function 232.

The domain discrimination loss function 234 can be expressed as: , where x is the original data. The processor 110 calculates the updating method of the weights of the domain discriminant model 144 and the feature extraction model 141 according to the domain discriminant loss function 234 and the domain adversarial training mechanism, so as to ensure that the abstract features extracted by the feature extraction model 141 have the characteristics of domain invariance.

The task loss function 233 may be expressed, for example, as: , where x is the original data, y* is the label, and L is the set of labeled data. The processor 110 uses the labeled samples to perform supervised learning of the task model and uses the back propagation algorithm to calculate the update direction of the weights of the task model 143 and the feature extraction model 141 to ensure that the model can make accurate judgments on the task.

The semi-supervisory loss function 232 may be expressed, for example, as: , where U is a set of unlabeled data, and D _KL is a metric for measuring two distributions (e.g., using the KL divergence algorithm). The processor 110 uses the unlabeled data to provide information for the feature extraction model 141 and the task model 143 to learn, and calculates the update direction of the weights of the semi-supervised learning mechanism 142 through the back propagation algorithm.

The processor 110 also assigns different weights to the domain discrimination loss function, the task loss function and the semi-supervision loss function, for example, , and Therefore, the processor 110 can calculate the weighted total loss function SL, as shown in formula (1):

Among them, the weight , and are all non-negative real numbers; x is the original data; y* is the label; D _KL is the measurement method for measuring the two distributions; L is the set of labeled data; U is the set of unlabeled data. , and Determined in advance by the user, usually weighted , and The sum of is 1. Its value directly reflects the impact of the corresponding loss function. Therefore, if you want to increase the importance of task loss, you can set a larger value. The value is also related to the number of data. If the number of labeled samples is very small, a larger value can be set. To avoid the loss function losing its influence.

The processor 110 also calculates the adjustment direction of the weights of each model in the overall model through the back propagation method according to the calculation result of the total loss function SL and uses it to update the weights of each model in the overall model. In detail, the processor 110 uses source domain data and target domain data from different source domains to calculate the domain discrimination loss function, and updates the weights of the feature extraction model and the weights of the domain discrimination model according to the domain discrimination loss function. The processor 110 uses the labeled data in the source domain data and the target domain data to calculate the task loss function, and updates the weights of the feature extraction model and the weights of the task model according to the task loss function. For example, the feature extraction model is pre-trained using available data, and common implementation methods include self-supervised learning or generative adversarial networks. When performing subsequent multi-source domain target task training, the weights of the feature extraction model will also be adjusted according to the update direction of the back propagation method calculated by the overall loss function, and at this time, it does not correspond to a dedicated loss function (Note: This means that the output of the feature extraction function is not used to calculate the total loss function SL). In addition, the processor 110 uses the unlabeled data in the source domain data and the target domain data to calculate the semi-supervised loss function, and updates the weights of the feature extraction model (and/or task model) based on the semi-supervised loss function. In some embodiments, depending on the semi-supervised learning mechanism 142 used by the processor 110, the processor 110 uses the labeled data and unlabeled data in the source domain data and the target domain data to calculate the semi-supervised loss function, and updates the weights of the feature extraction model (and/or task model) based on the semi-supervised loss function.

Then, the processor 110 repeatedly executes the steps of calculating the three loss functions and updating the model weights. When the feature extraction model 141 reaches the model convergence condition, the processor 110 ends the training process of the feature extraction model 141 and obtains the trained feature extraction model 141. For example, the above-mentioned model convergence condition can be that the overall model reaches the target value in a fixed training batch (epoch) or a fixed training time or a total loss function, etc., but the present invention is not limited thereto.

3A-3C are schematic diagrams of source domain data and target domain data according to an embodiment of the present invention.

In the field of industrial production, defect detection is a common quality inspection method to ensure that there are no defects in the products during the processing. In the electronics industry, soldering is a typical processing procedure and plays an important role in the yield of printed circuit board (PCB) type products. Therefore, corresponding automatic optical inspection (AOI) inspection stations are often configured to perform defect detection on solder. However, the defect rate of mature manufacturers is usually very low, resulting in a small number of valid label samples and high acquisition costs. Furthermore, there are many types of printed circuit boards and the shapes and angles of each solder joint are different, resulting in insufficient accuracy of defect detection and requiring a large amount of manual re-judgment. Therefore, in the embodiment of Figures 3A-3C, a semi-supervised learning system 100 is used to solve the problems faced by defect detection.

For ease of explanation, in this embodiment, before training the feature extraction model 141, the semi-supervised learning system 100 first obtains different amounts of first source domain data, second source domain data, and target domain data, wherein the first source domain data and the second source domain data are obtained from factory A and factory B, respectively, and are detection point images of each detection point (e.g., solder points) of the printed circuit board of model 80. The target domain data is obtained from factory A, and is a detection point image of some detection points (e.g., solder points) of the printed circuit board of model 60. Figure 3A is one of the detection point images in the first source domain data, and Figure 3B is one of the detection point images in the second source domain data. Figure 3C is one of the detection point images in the target domain data.

It should be noted that the shooting conditions of the automatic optical inspection equipment in Factory A and Factory B may be different. For example, they may be affected by light, shooting distance, shooting angle, and exposure time. In addition, there are many types of printed circuit boards and the shapes and angles of each solder joint are different, which leads to different inspection point images taken for the same inspection point of the same model of printed circuit boards.

In addition, the goal of the semi-supervised learning system 100 is to use the first source domain data and the second source domain data to import the defect detection function of the printed circuit board (i.e., the target domain) produced in Factory A Model 60, wherein the defect detection is a binary classification, which means that the judgment result of the sample (i.e., the inspection point image) is passed or failed. If the judgment result is passed, it means that the semi-supervised learning system 100 judges that the sample has no defects; if the judgment result is failed, it means that the semi-supervised learning system 100 judges that the sample has defects.

The number of good samples, defective samples, unlabeled samples and total samples in the first source domain data, the second source domain data and the target domain data are shown in Table 1: First source domain data Second source domain data Target domain data Number of good samples 26 493 17 Number of defective samples 138 7 41 Number of unlabeled samples 0 1504 0 Total number of samples 164 2004 58 Table 1

After the processor 110 inputs the first source domain data (for example, including 164 labeled data), the second source domain data (for example, including 500 labeled data and 1504 unlabeled data) and the target domain data (for example, 20 unlabeled data) into the feature extraction model 141 and completes the training through the process of the above embodiment, the processor 110 can use three indicators: accuracy, precision and recall to measure the classification ability of the feature extraction model 141.

For example, the number of defective products detected can be divided into four categories: true positive, true negative, false positive and false negative, and the corresponding numbers are represented by TP, TN, FP and FN respectively. TP represents the number of samples that the feature extraction model 141 judges to be defective and are actually defective. TN represents the number of samples that the feature extraction model 141 judges to be good and are actually good. FP represents the number of samples that the feature extraction model 141 mistakenly judges to be good as defective. FP represents the number of samples that the feature extraction model 141 mistakenly judges to be defective as good. The above representation method is for defect detection of defective products. A person skilled in the art can also use a similar method to calculate the four test quantities of good products.

For defective products, the accuracy of the feature extraction model 141 = (TP+TN)/ (TP+FP+TN+FN), which means the overall accuracy of the feature extraction model 141 in making correct judgments. The recall of the feature extraction model 141 = TP/(TP+FN), which means the proportion of defective products that the feature extraction model 141 can correctly judge as defective products among all defective products. The precision of the feature extraction model 141 = TP/(TP+FP), which means the proportion of samples judged as defective products by the feature extraction model 141 are actually defective products.

In a similar manner, the processor 110 can also calculate the accuracy, recall and precision of good products, where the accuracy of good products is the same as the accuracy of bad products, both referring to the overall accuracy of the correct judgment of the feature extraction model 141. The recall rate of good products indicates the percentage of good products that the feature extraction model 141 can correctly judge among all good products. The precision of good products indicates the percentage of samples judged as good products by the feature extraction model 141 that are truly good products.

Therefore, the accuracy, recall and precision for good and bad products are shown in Table 2: Tags Recall Accuracy Accuracy Good quality Good product recall rate Good product accuracy Overall accuracy Defective products Defective product recall rate Defective product accuracy rate Table 2

Here, the present invention provides two common machine learning methods for comparison with the performance of the semi-supervised learning system 100 of the present invention. Control group 1 adopts a typical supervised learning method and only uses 47 labeled data in the target domain data for training, and uses the remaining 11 labeled data for testing. The scenario of control group 1 represents the accuracy, recall and precision of the machine learning model without referring to similar data in other source domains by re-collecting multiple data in the target domain and consuming resources for labeling.

The accuracy, recall and precision of control group 1 for good and defective products are shown in Table 3: Tags Recall Accuracy Accuracy Good quality 0.94 0.94 0.92 Defective products 0.92 0.82 table 3

Control group 2 uses domain adaptation technology, and the training data of the machine learning model uses 164 labeled data in the first source domain data, 500 labeled data and 1504 unlabeled data in the second source domain data, and 20 unlabeled data in the target domain data. The scenario of control group 2 represents the performance of the machine learning model in the target domain when there are multiple source domains for reference and both labeled and unlabeled data, and only a small amount of unlabeled data in the target domain.

After the training of the machine learning model of control group 2 is completed, the remaining 38 data in the target domain data are used to test the machine learning model. The accuracy, recall and precision of control group 2 for good and defective products can be obtained, as shown in Table 4: Tags Recall Accuracy Accuracy Good quality 0.54 0.58 0.71 Defective products 0.80 0.77 Table 4

The semi-supervised learning system 100 of the present invention uses the same training data as the control group 2. After the feature extraction model 141 is trained through the process of the above embodiment, the processor 110 uses the remaining 38 data in the target domain data to test the machine learning model, and the accuracy, recall rate and precision rate of the semi-supervised learning system 100 for good and bad products can be obtained, as shown in Table 5: Tags Recall Accuracy Accuracy Good quality 1.00 0.81 0.92 Defective products 0.88 1.00 table 5

The scenario of the semi-supervised learning system 100 of the present invention indicates that it can fully utilize the labeled data and unlabeled data in the source domain data at the same time, and use a small amount of unlabeled data in the target domain for model transfer, which can achieve higher accuracy and precision than other methods, and can even approach the performance of the control group 1 when using a large amount of labeled data to establish a machine learning model, and has extremely high accuracy.

FIG. 4 is a flow chart of a semi-supervised learning method according to an embodiment of the present invention.

In step S410, source domain data of one or more source domains and target domain data of a target domain are obtained. For example, the source domain data of each source domain and the target domain data of the target domain both include labeled data and unlabeled data.

In step S420, the source domain data and the target domain data are used to train a feature extraction model. For example, in the embodiment of FIGS. 3A-3C, the processor 110 inputs the first source domain data (e.g., including 164 labeled data), the second source domain data (e.g., including 500 labeled data and 1504 unlabeled data) and the target domain data (e.g., 20 unlabeled data) into the feature extraction model 141 for training.

In step S430, a domain discrimination loss function is calculated. The domain discrimination loss function can be expressed as: , where x is the original data. The processor 110 calculates the updating method of the weights of the domain discriminant model 144 and the feature extraction model 141 according to the domain discriminant loss function and the domain adversarial training mechanism, so as to ensure that the abstract features extracted by the feature extraction model 141 have the characteristics of domain invariance.

In step S440, the task loss function is calculated. The task loss function can be expressed as: , where x is the original data, y* is the label, and L is the set of labeled data. The processor 110 uses the labeled samples to perform supervised learning of the task model and uses the back propagation algorithm to calculate the update direction of the weights of the task model 143 and the feature extraction model 141 to ensure that the model can make accurate judgments on the task.

In step S450, a semi-supervisory loss function is calculated. The semi-supervisory loss function can be expressed as: , where U is a set of unlabeled data, and D _KL is a metric for measuring two distributions (e.g., using a KL divergence algorithm). The processor 110 uses the unlabeled data to provide information for the feature extraction model 141 and the task model 143 to learn, and calculates the update direction of the weight of the feature extraction model 141 through a back propagation algorithm. In this embodiment, steps S430-450 can be integrated into the same step, or their order can be arbitrarily changed, or steps S430-450 can be executed simultaneously.

In step S460, the model weights are updated. For example, the processor 110 may calculate and update the first weight, the second weight, and the third weight corresponding to the feature extraction model, the task model, and the domain discrimination model in the overall model through the back propagation method according to the calculation result of the total loss function SL.

In step S470, it is determined whether the overall model satisfies the model convergence condition. If so, step S480 is executed to end the training process of the overall model. If not, the process returns to step S430 to repeat the steps of calculating the three loss functions and updating the model weights. For example, the above-mentioned model convergence condition may be that the overall model reaches the target value in a fixed training batch (epoch) or a fixed training time or a total loss function, etc., but the present invention is not limited thereto.

In summary, the present invention provides a semi-supervised learning system and a semi-supervised learning method, which can combine domain adaptation technology with a semi-supervised learning mechanism, and can use source domain data and target domain data for model training, and can be combined with a domain discrimination loss function, a task loss function and a semi-supervised loss function to update the model weights, so that the model has a better training stage to achieve better machine learning performance.

The method of the present invention, or a specific form or part thereof, may be included in the form of program code on a physical medium, such as a floppy disk, an optical disk, a hard disk, or any other machine-readable (such as computer-readable) storage medium, wherein when the program code is loaded and executed by a machine, such as a computer, the machine becomes a device or system for participating in the present invention. The method, system and device of the present invention may also be transmitted in the form of program code through some transmission medium, such as wires or cables, optical fibers, or any transmission type, wherein when the program code is received, loaded and executed by a machine, such as a computer, the machine becomes a device or system for participating in the present invention. When implemented on a general-purpose processor, the program code combines with the processor to provide a unique device that operates similarly to an application-specific logic circuit.

Although the present invention is disclosed as above with the preferred embodiments, it is not intended to limit the scope of the present invention. Any person with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be defined by the scope of the attached patent application.

100: Semi-supervised learning system 110: Processor 130: Memory unit 140: Storage device 141: Feature extraction model 142: Semi-supervised learning mechanism 143: Task model 144: Domain discrimination model 210-21N: Source domain data 220: Target domain data 210A-21NA, 220A: Labeled data 210B-21NB, 220B: Unlabeled data 231: Feature vector 232: Semi-supervised loss function 233: Task loss function 234: Domain discrimination loss function S410-S480: Steps

FIG. 1 is a schematic diagram of a semi-supervised learning system according to an embodiment of the present invention. FIG. 2 is a schematic diagram of a training process of a feature extraction model according to an embodiment of the present invention. FIG. 3A-3C are schematic diagrams of source domain data and target domain data according to an embodiment of the present invention. FIG. 4 is a flow chart of a semi-supervised learning method according to an embodiment of the present invention.

S410-S480: Steps

Claims (10)

A semi-supervised learning system includes: a non-volatile memory for storing a semi-supervised learning application; and a processor for executing the semi-supervised learning application to perform the following steps: obtaining source domain data of one or more source domains and target domain data of a target domain; using the source domain data and the target domain data to train a feature extraction model; using a domain discrimination model, a task The invention relates to a method for calculating a domain discrimination loss function, a task loss function and a semi-supervised loss function of the feature extraction model respectively by using a model and a semi-supervised learning mechanism; a total loss function is calculated according to the domain discrimination loss function, the task loss function and the semi-supervised loss function, and a first weight, a second weight and a second weight corresponding to the feature extraction model, the task model and the domain discrimination model are updated according to the total loss function. weight and a third weight; and in response to an overall model satisfying a model convergence condition, the training process of the overall model is terminated, wherein the overall model includes the feature extraction model, the task model and the domain discrimination model; wherein the source domain data includes first labeled data and first unlabeled data, and the target domain data includes second labeled data and second unlabeled data; wherein the processor uses the first labeled data, the first unlabeled data, the second labeled data and the second unlabeled data to calculate the domain discrimination loss function; wherein the processor uses the first labeled data and the second labeled data to calculate the task loss function; and wherein the processor uses the first unlabeled data and the second unlabeled data to calculate the semi-supervised loss function. The semi-supervised learning system of claim 1, wherein the feature extraction model is a ResNet50 model, the semi-supervised learning mechanism is a consistency-normalized unsupervised data augmentation mechanism, the task model is a first fully connected layer, and the domain discrimination model is a second fully connected layer plus a gradient inversion layer. A semi-supervised learning system as claimed in claim 1, wherein the processor assigns the first hyperparameter, the second hyperparameter and the third hyperparameter to the domain judgment loss function, the task loss function and the semi-supervised loss function respectively to calculate the total loss function. A semi-supervised learning system as claimed in claim 3, wherein the processor updates the feature extraction model and the first weight of the domain discrimination model according to the domain discrimination loss function, updates the first weight of the feature extraction model and the second weight of the task model according to the task loss function, and updates the first weight of the feature extraction model according to the semi-supervised loss function. As in claim 2, the semi-supervised learning system, wherein the model convergence condition means that the overall model reaches the target value in a fixed training batch (epoch) or a fixed training time or the total loss function. A semi-supervised learning method includes: obtaining source domain data of one or more source domains and target domain data of a target domain; using the source domain data and the target domain data to train a feature extraction model; using a domain discrimination model, a task model and a semi-supervised learning mechanism to respectively calculate the domain discrimination loss function, the task loss function and the semi-supervised loss function of the feature extraction model; calculating a total loss function based on the domain discrimination loss function, the task loss function and the semi-supervised loss function, and updating the first weight, the second weight and the third weight corresponding to the feature extraction model, the task model and the domain discrimination model based on the total loss function; and updating the first weight, the second weight and the third weight corresponding to the feature extraction model, the task model and the domain discrimination model in response to the fullness of an overall model. The model convergence condition is satisfied, and the training process of the overall model is terminated, wherein the overall model includes the feature extraction model, the task model and the domain discrimination model; wherein the source domain data includes the first labeled data and the first unlabeled data, and the target domain data includes the second labeled data and the second unlabeled data, and the method further includes: using the first labeled data, the first unlabeled data, the second labeled data and the second unlabeled data to calculate the domain discrimination loss function; using the first labeled data and the second labeled data to calculate the task loss function; and using the first unlabeled data and the second unlabeled data to calculate the semi-supervised loss function. The semi-supervised learning method of claim 6, wherein the feature extraction model is a ResNet50 model, the semi-supervised learning mechanism is a consistency regularized unsupervised data augmentation mechanism, the task model is a first fully connected layer, and the domain discrimination model is a second fully connected layer plus a gradient inversion layer. The semi-supervised learning method of claim 7 further includes: respectively assigning the first hyperparameter, the second hyperparameter and the third hyperparameter to the domain judgment loss function, the task loss function and the semi-supervised loss function to calculate the total loss function. The semi-supervised learning method of claim 8 further includes: updating the first weight of the feature extraction model according to the domain discrimination loss function; updating the first weight of the feature extraction model and the second weight of the task model according to the task loss function; and updating the first weight of the feature extraction model according to the semi-supervised loss function. As in claim 7, the semi-supervised learning method, wherein the model convergence condition means that the overall model reaches the target value in a fixed training batch (epoch) or a fixed training time or the total loss function.

Images