Sample knowledge acquisition and continuous utilization method

Abstract

The present invention discloses a method for acquiring and continuously utilizing sample knowledge. The method comprises: first, obtaining an example sample set from an existing labeled sample set; based on the example sample set, obtaining its corresponding feature extraction model and the task model for each task, and saving them to a knowledge base; then, comparing the new task with the tasks in the knowledge base through task similarity calculation, obtaining the task with the highest similarity to the new task, and utilizing the knowledge content of the task to serve the new task. The present invention can utilize the knowledge content of existing samples and fully utilize this prior knowledge to achieve continuous knowledge utilization and task learning, reducing the investment of manpower and material resources by constructing a learning model.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing sample-related knowledge acquisition and application, and in particular to a method for sample knowledge acquisition and continuous utilization. Background Art

[0002] Although deep neural networks have become one of the most successful deep learning technologies, with recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and convolutional neural networks (CNNs) achieving impressive performance, the current mainstream deep learning paradigm remains isolated learning, which disregards other relevant information and previously learned knowledge. The main problem with this isolated learning approach is that it fails to retain or accumulate previously learned knowledge, making it impossible to leverage it in future learning. This stands in stark contrast to the human learning process.

[0003] Humans never learn in isolation or from scratch. We always retain past knowledge and use it to inform future learning and problem-solving. Without the ability to accumulate and utilize previously acquired knowledge, deep learning algorithms often require a large number of training samples to effectively learn, a typically static and closed learning environment. For most neural network learning tasks, the success of the learning task primarily stems from the complexity of the network structure and the richness of the training samples. Training samples, in particular, typically require a large number of labeled samples from multiple categories, which is a very time-consuming and costly task. Furthermore, since things are not static, labeling work must be ongoing and cannot be done once and for all.

[0004] In the field of surveying and mapping and remote sensing, with the continuous development of aerospace technology and communication technology, high-resolution and hyperspectral satellite images are becoming more and more abundant. We can obtain more and more abundant satellite remote sensing data, thereby realizing high-precision surface monitoring over a large area, thus serving agriculture, geology, water conservancy, environment and other fields.

[0005] However, due to factors such as atmospheric absorption and scattering, sensor calibration, solar altitude, azimuth, phenological phase, and data processing, the spectral data of ground objects acquired by satellite sensors often changes over time. This results in remote sensing images often showing different objects with the same spectrum, or different objects with the same spectrum. Furthermore, remote sensing images taken from the same region at different times may have spectral values ​​that do not follow the same probability statistical distribution. This makes sample collection and model training tasks more complex and arduous than in other fields. Therefore, how to leverage the "prior knowledge" of existing samples and models trained based on these samples to fully utilize past knowledge is an urgent problem that needs to be solved. Summary of the Invention

[0006] In response to the problem of insufficient utilization of a large amount of prior information and knowledge in the current remote sensing imagery for ground object extraction, the present invention constructs a method for acquiring and continuously utilizing sample knowledge, which fully utilizes the prior information and knowledge of existing samples, reduces the investment of manpower and material resources, realizes continuous knowledge learning, and improves the classification robustness of the current remote sensing image classification task.

[0007] The present invention provides a method for acquiring and continuously utilizing sample knowledge, the method comprising:

[0008] S1 obtains an existing labeled sample set, and constructs a sample set based on the labeled sample set;

[0009] S2 constructs a first task using the example sample set to obtain a first task feature extraction model and a first task model;

[0010] S3 combines the first task, the example sample set, the first task feature extraction model, and the first task model into a knowledge sequence, and saves the knowledge sequence into a knowledge base;

[0011] S4 repeats S1-S3 to enrich the knowledge base, so that the knowledge base includes N knowledge sequences, where N ≥ 1;

[0012] S5: Given a second task and labeled sample data of the second task, using the labeled sample data of the second task, calculating the relevance between the second task and each task in the knowledge base, and selecting the task with the greatest relevance from the knowledge base as a target task related to the second task;

[0013] S6: Migrate the task model of the target task related to the second task according to the value of the task relevance.

[0014] In a specific embodiment of the present invention, constructing a sample set based on the labeled sample set includes:

[0015] Constructing a first feature extraction model based on the labeled sample set;

[0016] Using the first feature extraction model, extracting features from the labeled sample set to obtain a sample feature vector of the labeled sample set;

[0017] According to the sample feature vectors of the labeled sample set and the similarity of the sample features, a sample set of examples of the labeled sample set is obtained.

[0018] In a specific embodiment of the present invention, the first feature extraction model is trained using a neural network method.

[0019] In a specific embodiment of the present invention, obtaining the exemplary sample set of the labeled sample set according to the sample feature vector of the labeled sample set and the similarity of the sample features includes:

[0020] Calculating an average feature vector corresponding to the labeled sample set according to the sample feature vectors of the labeled sample set;

[0021] Select m samples from the labeled sample set, wherein the difference between the average sample feature vector of the m samples and the average feature vector corresponding to the labeled sample set is the smallest, where m ≥ 1 and is less than the number of samples in the labeled sample set;

[0022] The m samples are sorted in ascending order according to the difference between the m samples and the average sample feature vector of the m samples, so as to obtain a sample set of examples of the labeled sample set.

[0023] In a specific embodiment of the present invention, the first task, the example sample set, the first task feature extraction model, and the first task model are combined into a knowledge sequence, and the knowledge sequence is saved in a knowledge base, including:

[0024] Constructing a tuple according to the first task, the example sample set, the first task feature extraction model, and the first task model;

[0025] The tuple is used as a knowledge sequence, and the knowledge sequence is saved in a knowledge base.

[0026] In a specific embodiment of the present invention, the first task feature extraction model adopts an undercomplete autoencoder, which is composed of an encoder and a decoder, including an input layer, a hidden layer and an output layer, and a ReLU layer is provided between each hidden layer.

[0027] In a specific embodiment of the present invention, the calculating the relevance between the second task and each task in the knowledge base using the labeled sample data of the second task includes:

[0028] Select the kth task T from the knowledge base k , get T k The corresponding example sample set X k , wherein the example sample set X k By sample Composition, where i = 1, 2, 3, ..., n;

[0029] Calculate the samples in the knowledge base The reconstruction error applied to the undercomplete autoencoder is calculated based on the sample The reconstruction error applied to the incomplete autoencoder is used to obtain the example sample set X ka reconstruction error applied to the undercomplete autoencoder;

[0030] Obtain the example sample set X′ corresponding to the second task, wherein the example sample set X′ consists of sample x′ j Composition, where j = 1, 2, 3, ..., m;

[0031] Calculate the sample x' in the second task j The reconstruction error applied to the undercomplete autoencoder is based on the sample x′ in the second task j Applying the reconstruction error of the undercomplete autoencoder to obtain the reconstruction error of the example sample set X′ corresponding to the second task applied to the undercomplete autoencoder;

[0032] According to the example sample set X k The reconstruction error applied to the incomplete autoencoder and the reconstruction error of the example sample set X′ applied to the incomplete autoencoder are used to obtain the second task and T k Task relevance.

[0033] In a specific embodiment of the present invention, migrating the task model of the target task related to the second task according to the value of the task relevance includes:

[0034] When the value of the task relevance is greater than or equal to a preset relevance threshold, migrating the task model of the target task related to the second task by using a shared model;

[0035] When the value of the task relevance is less than a preset relevance threshold, a task model of the target task related to the second task is migrated in a fine-tuning manner.

[0036] In a specific embodiment of the present invention, migrating the task model of the target task related to the second task by using a shared model includes:

[0037] Obtaining parameters shared by the second task and the target task, parameters learned in the target task, and randomly initialized parameters in the second task;

[0038] Model migration is performed according to parameters shared by the second task and the target task, parameters learned in the target task, and randomly initialized parameters in the second task.

[0039] In a specific embodiment of the present invention, the step of migrating the task model of the target task related to the second task by fine-tuning includes:

[0040] Locking a preset network layer in a task model of the target task;

[0041] The locked task model is updated and iterated according to the labeled sample data of the second task to achieve model migration.

[0042] The present invention provides a method for acquiring and continuously utilizing sample knowledge. First, a sample set is obtained from an existing labeled sample set. A feature extraction model and a task model for each task are obtained using the sample set. The sample set, the feature extraction model of the sample set, and the task model for each task are saved. Then, a new task is compared with an existing task using a task similarity calculation method. The existing task in the knowledge base with the highest similarity to the new task is obtained, and the knowledge content corresponding to the existing task is obtained to serve the new task. The beneficial effects of the present invention are:

[0043] (1) Compared with traditional machine learning and deep learning models, the patent of this invention provides a beneficial continuous knowledge learning method, which fully utilizes the components of prior knowledge and the method content of continuous learning, guides and constructs a learning model close to human thinking, and reduces the investment of manpower and material resources.

[0044] (2) Aiming at the situation where efficient storage and efficient transfer of knowledge are required in the continuous utilization and learning of samples, the present invention adopts the average eigenvector method to obtain typical samples, ensuring that the general properties of all samples are not reduced when some samples are removed; the sample set reconstruction error is used to obtain the most similar prior knowledge in the knowledge base, and the knowledge transfer strategy is unified according to the task relevance, reducing the blind obedience of knowledge transfer and ensuring the framework consistency and method efficiency of the continuous utilization of sample knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 A flowchart of an embodiment of a method for acquiring and continuously utilizing sample knowledge provided by the present invention;

[0047] Figure 2 A model architecture of an undercomplete autoencoder established for an embodiment of the method for acquiring and continuously utilizing sample knowledge of the present invention;

[0048] Figure 3 A model architecture of a model used by the first task model of an embodiment of the method for acquiring and continuously utilizing sample knowledge of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention are within the scope of protection of the present invention.

[0050] Figure 1 The figure shows a flow chart of a method for acquiring and continuously utilizing sample knowledge provided by the present invention. Compared with traditional machine learning and deep learning models, the present invention provides a beneficial continuous knowledge learning method that fully utilizes the components of prior knowledge and the method content of continuous learning, guides and constructs a learning model close to human thinking, and reduces the investment of manpower and material resources. Specifically, the method includes the following steps:

[0051] S1 obtains an existing labeled sample set, and constructs a sample set based on the labeled sample set;

[0052] (1) Obtain an existing set of labeled samples

[0053] Deep learning is mainly used in the image field for image classification, target detection, and semantic segmentation. Image classification is to classify the image type based on the object information in the image. Target detection is to find the detected objects in the image and frame their positions using the object bounding box. Semantic segmentation is to find the detected objects in the image and accurately extract the boundaries of the objects.

[0054] The existing labeled sample sets include samples for image classification, object detection, and semantic segmentation. In this example, a labeled sample set for image classification is selected and denoted as X0.

[0055] (2) Preprocessing the labeled sample set

[0056] The sample size is scaled and all samples in the labeled sample set are unified into a scale of 224×224. The sample set is denoted as in For the sample set In one sample, There are n samples in total.

[0057] (3) Construct the first feature extraction model

[0058] Adopting the AlexNet model architecture, inputting a sample set with uniform scale The FC7 layer after model training is used as the output to obtain the first feature extraction model The output of is a feature vector.

[0059] (4) using the first feature extraction model to extract features from the labeled sample set to obtain a sample feature vector of the labeled sample set;

[0060] for Extract model using the first feature Solution The eigenvectors of:

[0061]

[0062] where f i is x i The eigenvector of is a set of feature vectors.

[0063] (5) Based on the labeled sample set The sample feature vector of Obtaining a sample set of examples of the labeled sample set according to the similarity of the sample features;

[0064] 1) Calculate the average feature vector of the labeled sample set based on the sample feature vectors of the labeled sample set:

[0065]

[0066] Where, is the average eigenvector.

[0067] 2) Select m samples from the labeled sample set, and the difference between the average sample feature vector of the m samples and the average feature vector corresponding to the labeled sample set is the smallest, where m ≥ 1 and is less than the number of samples in the labeled sample set, that is:

[0068]

[0069] Where X k The sample is selected.

[0070] The above steps use the average eigenvector method to obtain typical samples, ensuring that the general properties of all samples are not reduced when some samples are removed.

[0071] (3) Sort the m samples in ascending order according to the difference between the m samples and the average sample feature vector, and obtain the example sample set of the labeled sample set, and denote the example sample set as X, X = {x1, x2, ..., x k}.

[0072] S2 constructs a first task using the example sample set to obtain a first task feature extraction model and a first task model;

[0073] In this example, the first task is an image classification task, denoted as T1; the feature extraction model of the first task uses an undercomplete autoencoder, denoted as A1; the first task model is composed of DLinkNet, and its input is the features extracted by A1, denoted as E1.

[0074] An undercomplete autoencoder is one in which the encoding dimension is smaller than the decoding dimension. An undercomplete autoencoder consists of an encoder and a decoder. The encoder reshapes the original data, preserving as much useful information as possible while removing or minimizing useless information. The decoder uses the encoder's encoding output to decrypt the original information.

[0075] The goal of learning an undercomplete autoencoder is to make the original data and the output data as similar as possible. This process forces the autoencoder to capture the most significant features of the input data. In simple terms, an undercomplete autoencoder learns a low-dimensional feature representation that best describes the data in the most compact way.

[0076] like Figure 2 As shown, the incomplete autoencoder A1 consists of an encoder and a decoder, including an input layer, a hidden layer, and an output layer. k} is used as the input of the input layer. The data in the hidden layer represents the feature data encoded by the network training. The data in the output layer is the decoded data reconstructed from the feature data. The input dimension of the incomplete autoencoder A1 is 224×224×3. The encoder constructs 3 hidden layers to compress the data dimension to 256. There is a ReLU layer between each hidden layer to speed up the training and overcome the problem of gradient disappearance. The decoder decodes the data dimension from 256 to 224×224×3. There is a ReLU layer between each hidden layer of the decoder. There is a Tanh layer between the last hidden layer and the output layer, and features are transferred through the Tanh activation function. The decoder takes the output of the encoder as input and finally outputs X'={x1′,x2′,...,x k ′}, so that x i ′ as close as possible to x i Similar, where 1≤i≤k.

[0077] The first task model uses VGGNet11, whose structure is as follows Figure 3 Shown, including:

[0078] (1) Input layer: Input 224×224×3 sample data;

[0079] (2) Perform 3×3 convolution + Rule, and the size after convolution becomes 224×224×64;

[0080] (3) Perform 2×2 max pooling, and the size becomes 112×112×64. Then, perform one convolution + ReLU with 128 3×3 convolution kernels, and the size becomes 112×112×128.

[0081] (4) Perform 2×2 max pooling, and the size becomes 56×56×128. Then, perform two convolutions + ReLU with 256 3×3 convolution kernels, and the size becomes 56×56×256.

[0082] (5) Perform 2×2 max pooling, and the size becomes 28×28×256. Then, perform two convolutions + ReLU with 512 3×3 convolution kernels, and the size becomes 28×28×512.

[0083] (6) Perform 2×2 max pooling, and the size becomes 14×14×512. Then, perform two convolutions + ReLU with 512 3×3 convolution kernels, and the size becomes 14×14×512.

[0084] (7) Perform 2×2 max pooling (maximum pooling), and the size becomes 7×7×512;

[0085] (8) Fully connect with two layers of 1×1×4096 and one layer of 1×1×15 + ReLU (three layers in total), and output 15 prediction results through softmax.

[0086] S3 combines the first task, the example sample set, the first task feature extraction model, and the first task model into a knowledge sequence, and saves the knowledge sequence into a knowledge base;

[0087] In this example, the first task T1, the example sample set X, the first task feature extraction model A1, and the first task model E1 are constructed to form a tuple {T1, X, A1, E1} as the knowledge sequence of the first task T1, and a non-relational database is used as the knowledge base to store the tuple.

[0088] S4 repeats S1-S5 to enrich the knowledge base so that the knowledge base includes N knowledge sequences, where N ≥ 1;

[0089] S5: Given a second task and labeled sample data of the second task, using the labeled sample data of the second task, calculating the relevance between the second task and each task in the knowledge base, and selecting the task with the greatest relevance from the knowledge base as a target task related to the second task;

[0090] In this example, the second task is denoted as T, and the labeled sample data of the second task is denoted as X. t :

[0091] (1) Select the kth task T from the knowledge base k , get T k The corresponding example sample set X k , where the example sample set X k From the sample x i Composition, where i = 1, 2, 3, ... n;

[0092] (2) Calculate sample x i The reconstruction error applied to the undercomplete autoencoder A1 is:

[0093] er i =L(x i ,g(f(x i )))(4)

[0094] Among them, f(x i ) indicates that the encoder function converts x i Mapping encoder space, g(f(x i ))Through the decoder function, f(x i ) to restore, er i Indicates that x i The same result after encoding and decoding g(f(x i ))’s mean square error.

[0095] (3) Calculate the example sample set X k The reconstruction error applied to the undercomplete autoencoder A1 is:

[0096]

[0097] in, Represents X k The reconstruction error of the sample x in |X k | for X k The number of samples.

[0098] (4) Calculate the reconstruction error of the second task based on the reconstruction error of the samples in the second task applied to the undercomplete autoencoder:

[0099]

[0100] Among them, Er t is the reconstruction error of the second task, Represents X t The reconstruction error of the sample x in |X t | for Xt The number of samples.

[0101] (5) Calculate the second task and T k The calculation formula of the task relevance is:

[0102]

[0103] Wherein, T is the second task.

[0104] S6: Migrate the task model of the target task related to the second task according to the value of the task relevance.

[0105] In this example, when the value of the task relevance is greater than or equal to the preset relevance threshold, the task model of the target task related to the second task is migrated using a shared model; when the value of the task relevance is less than the preset relevance threshold, the task model of the target task related to the second task is migrated using a fine-tuning method. The preset relevance threshold may be 0.85 or other parameters, which are not limited in this embodiment. The target task related to the second task is the task related to the second task T. For example, the task most related to the second task T is recorded as T0, and the task model of the task most related to the second task T is recorded as E0. The numerical threshold of the relevance is set to 0.85. If the relevance value is greater than or equal to 0.85, it means that the second task T is sufficiently correlated with T0, and the shared model-based method is used to migrate E0; if the relevance value is less than 0.85, the fine-tuning-based method is used to migrate E0.

[0106] (1) Model migration based on shared models

[0107] In this embodiment, parameters shared by the second task and the target task, parameters learned in the target task, and randomly initialized parameters in the second task are obtained;

[0108] Model migration is performed according to parameters shared by the second task and the target task, parameters learned in the target task, and randomly initialized parameters in the second task.

[0109] In the specific implementation, the model migration is completed through the following steps:

[0110] 1) Note θ s is a set of parameters shared by the second task T and T0, θ0 is a set of parameters learned specifically for T0, θ n Parameters randomly initialized for the second task T.

[0111] 2) Initialization settings

[0112] Using θ s , θ0, labeled sample data X of the second task T t , using the E0 model to get the predicted value Y0, recorded as:

[0113] Y0=CNN E0 (X t ,θ s ,θ0)

[0114] Initialize θ with random weights n , recorded as:

[0115] θ n ←RandInit(|θ n |)

[0116] 3) Iterative training to obtain updated parameters

[0117]

[0118] in:

[0119] Y′ n To use θ s ,θ n The current parameter θ' s ,θ' n Get the predicted value;

[0120] Y′ n To use θ s ,θ n The current parameter θ' s ,θ' n Get the predicted value;

[0121] Y′0 is the value of θ s 、The current parameter θ' of θ0 s , θ'0 to get the predicted value;

[0122] L new (Y n ,Y′ n ) is the predicted value Y′ n and the true value Y n The loss is recorded as:

[0123] L new (Y n ,Y′ n )=-Y n logY′ n

[0124] L old(Y0,Y′0) is the loss of the predicted value Y′0 and the recorded value Y0, which is expressed as:

[0125]

[0126] Where l is the number of labeled samples, and To correct the probability:

[0127]

[0128] γ(θ s ,θ0,θ n ) is a regularization term used to avoid overfitting.

[0129] (2) Model transfer based on fine-tuning

[0130] In this embodiment, model migration is achieved by locking the preset model layer in the task model of the target task; and updating and iterating the task model of the target task according to the labeled sample data of the second task, wherein the preset model layer is the network layer before the fully connected layer.

[0131] In the specific implementation, the model migration is completed through the following steps:

[0132] 1) Lock all parameters before the fully connected layer of model E, which do not participate in gradient updates;

[0133] 2) Input the labeled sample data of the second task into model E and only update the parameters of the fully connected layer;

[0134] 3) When the loss of the fully connected layer stops decreasing for n consecutive times, all parameters are released and trained together to obtain a fine-tuned model.

[0135] The above steps of selecting the most relevant tasks and model migration from the knowledge base use the sample set reconstruction error to obtain the most similar prior knowledge in the knowledge base, and unify the knowledge migration strategy according to the task relevance to reduce the blind obedience of knowledge migration and ensure the framework consistency and method efficiency of the continuous utilization of sample knowledge.

Claims

1. A method for acquiring and continuously utilizing sample knowledge, characterized in that: The method includes: S1 obtains an existing labeled sample set, and constructs a paradigm sample set based on the labeled sample set, wherein the labeled sample set includes samples for image classification, image object detection, or image semantic segmentation; S2 constructs a first task using the example sample set to obtain a first task feature extraction model and a first task model; S3 combines the first task, the example sample set, the first task feature extraction model, and the first task model into a knowledge sequence, and saves the knowledge sequence into a knowledge base; S4 repeats S1-S3 to enrich the knowledge base, so that the knowledge base includes N knowledge sequences, where N ≥ 1; S5: Given a second task and labeled sample data of the second task, using the labeled sample data of the second task, calculating the relevance between the second task and each task in the knowledge base, and selecting the task with the greatest relevance from the knowledge base as a target task related to the second task; S6, performing model migration on the task model of the target task related to the second task according to the value of the task relevance; The first task feature extraction model adopts an undercomplete autoencoder, which consists of an encoder and a decoder, including an input layer, a hidden layer and an output layer, and a ReLU layer is provided between each hidden layer; The calculating the relevance between the second task and each task in the knowledge base using the labeled sample data of the second task includes: Select the kth task Tk from the knowledge base and obtain the example sample set Xk corresponding to Tk, wherein the example sample set Xk consists of sample Composition, where i = 1, 2, 3, ..., n; Calculate the samples in the knowledge base The reconstruction error applied to the undercomplete autoencoder is calculated based on the sample Applying the reconstruction error of the undercomplete autoencoder to obtain the reconstruction error of the example sample set Xk applied to the undercomplete autoencoder; Obtaining a sample set X′ corresponding to the second task, wherein the sample set X′ consists of samples x′j, where j=1, 2, 3, ..., m; Calculating the reconstruction error of the sample x′j in the second task when applied to the undercomplete autoencoder, and obtaining the reconstruction error of the example sample set X′ corresponding to the second task when applied to the undercomplete autoencoder based on the reconstruction error of the sample x′j in the second task when applied to the undercomplete autoencoder; The correlation between the second task and the Tk task is obtained based on the reconstruction error of the example sample set Xk applied to the undercomplete autoencoder and the reconstruction error of the example sample set X′ applied to the undercomplete autoencoder.

2. The method for acquiring and continuously utilizing sample knowledge according to claim 1, characterized in that: The step of constructing a sample set according to the labeled sample set includes: Constructing a first feature extraction model based on the labeled sample set; Using the first feature extraction model, extracting features from the labeled sample set to obtain a sample feature vector of the labeled sample set; According to the sample feature vectors of the labeled sample set and the similarity of the sample features, a sample set of examples of the labeled sample set is obtained.

3. The method for acquiring and continuously utilizing sample knowledge according to claim 2, characterized in that: The first feature extraction model is trained using a neural network method.

4. The method for acquiring and continuously utilizing sample knowledge according to claim 2, characterized in that: The step of obtaining, based on the sample feature vectors of the labeled sample set and the similarity of the sample features, a sample set of examples of the labeled sample set includes: Calculating an average feature vector corresponding to the labeled sample set according to the sample feature vectors of the labeled sample set; Select m samples from the labeled sample set, wherein the difference between the average sample feature vector of the m samples and the average feature vector corresponding to the labeled sample set is the smallest, where m ≥ 1 and is less than the number of samples in the labeled sample set; The m samples are sorted in ascending order according to the difference between the m samples and the average sample feature vector of the m samples, so as to obtain a sample set of examples of the labeled sample set.

5. The method for acquiring and continuously utilizing sample knowledge according to claim 1, characterized in that: The step of forming a knowledge sequence from the first task, the example sample set, the first task feature extraction model, and the first task model, and saving the knowledge sequence to a knowledge base includes: Constructing a tuple according to the first task, the example sample set, the first task feature extraction model, and the first task model; The tuple is used as a knowledge sequence, and the knowledge sequence is saved in a knowledge base.

6. The method for acquiring and continuously utilizing sample knowledge according to any one of claims 1 to 5, characterized in that: The step of migrating the task model of the target task related to the second task according to the value of the task relevance includes: When the value of the task relevance is greater than or equal to a preset relevance threshold, migrating the task model of the target task related to the second task by using a shared model; When the value of the task relevance is less than a preset relevance threshold, a task model of the target task related to the second task is migrated in a fine-tuning manner.

7. The method for acquiring and continuously utilizing sample knowledge according to claim 6, characterized in that: The step of migrating the task model of the target task related to the second task by adopting a shared model includes: Obtaining parameters shared by the second task and the target task, parameters learned in the target task, and randomly initialized parameters in the second task; Model migration is performed according to parameters shared by the second task and the target task, parameters learned in the target task, and randomly initialized parameters in the second task.

8. The method for acquiring and continuously utilizing sample knowledge according to claim 6, characterized in that: The fine-tuning method of migrating the task model of the target task related to the second task includes: Locking a preset network layer in a task model of the target task; The locked task model is updated and iterated according to the labeled sample data of the second task to achieve model migration.

Images