Method, electronic device and computer program product for generating an image sample

Abstract

Embodiments of the present disclosure provide a method, an electronic device and a computer program product for generating image samples. The method comprises obtaining a set of distilled samples by processing a set of image samples using a classification capsule network model. A feature distribution of the set of distilled samples is indicative of a feature distribution of the set of image samples. Based on labels of the set of image samples, soft labels of each distilled sample in the set of distilled samples are determined, the soft labels representing probabilities of the distilled sample belonging to each of a plurality of classifications. In this way, by proposing classification capsules in the trained capsule network model as distilled samples, and assigning soft labels to the distilled samples based on the original image samples, it is possible to ensure that the distilled samples extracted have a high level of interpretability.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computers, and more specifically, to methods for generating image samples, electronic devices, and computer program products. Background Technology

[0002] Today, AI industry users, such as autonomous driving companies, rely on massive data computing power. Managing data, machine learning models, and underlying IT systems is complex and expensive. Therefore, the expectation for training datasets is that training sets containing a large number of samples can be transformed into training sets containing only a small number of samples, while ensuring that the transformed training set achieves the same training effect as the original. A common approach to reducing the training set is to distill the dataset, resulting in a dataset with a very small number of samples. This distilled dataset can then replace the original dataset for machine learning training. Summary of the Invention

[0003] Embodiments of this disclosure provide a scheme for generating distillation image samples using a classification capsule network.

[0004] In a first aspect of this disclosure, a method for generating image samples is provided. The method includes processing a set of image samples using a classification capsule network model to obtain a set of distilled samples, the feature distribution of which indicates the feature distribution of the set of image samples. The method also includes determining a soft label for each distilled sample in the set of distilled samples based on the labels of the set of image samples, the soft label representing the probability that the distilled sample belongs to each of a plurality of categories.

[0005] In a second aspect of this disclosure, an electronic device is provided. The electronic device includes a processor and a memory coupled to the processor, the memory having instructions stored therein, the instructions causing the device to perform actions when executed by the processor. The actions include: processing a set of image samples using a classification capsule network model to obtain a set of distilled samples, the feature distribution of the set of distilled samples indicating the feature distribution of the set of image samples; and determining a soft label for each distilled sample in the set of distilled samples based on the labels of the set of image samples, the soft label representing the probability that the distilled sample belongs to each of a plurality of categories.

[0006] In a third aspect of this disclosure, a computer program product is provided, which is tangibly stored on a computer-readable medium and includes machine-executable instructions that, when executed, cause a machine to perform the method according to the first aspect.

[0007] The summary section is provided to present the chosen concepts in a simplified form, which will be further described in the detailed description below. The summary section is not intended to identify key or principal features of this disclosure, nor is it intended to limit the scope of this disclosure. Attached Figure Description

[0008] The above and other objects, features, and advantages of this disclosure will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts. In the drawings:

[0009] Figure 1 A schematic diagram of an example environment in which embodiments of the present disclosure may be implemented is shown;

[0010] Figure 2 A flowchart illustrating an example method for generating image samples according to embodiments of the present disclosure is shown;

[0011] Figure 3 A flowchart illustrating an example method for determining soft tags according to embodiments of the present disclosure is shown;

[0012] Figure 4 A schematic diagram illustrating an example process for generating image samples according to some embodiments of the present disclosure is shown; and

[0013] Figure 5 A block diagram of an example device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0014] The principles of this disclosure will now be described with reference to several exemplary embodiments illustrated in the accompanying drawings. While preferred embodiments of this disclosure are shown in the drawings, it should be understood that these embodiments are described only to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way.

[0015] The term "comprising" and its variations as used herein signify open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "at least partially based on". The terms "one example embodiment" and "one embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0016] As discussed above, data distillation is typically based on convolutional neural networks (CNNs). However, existing data distillation schemes still have some drawbacks. For example, due to the limitations of CNN algorithms, the reliability of the labels on the distilled data is not high, resulting in poor interpretability and robustness. Therefore, it is desirable to find more optimized data distillation schemes to improve the quality of distilled data and its corresponding labels.

[0017] Embodiments of this disclosure propose a scheme for allocating tasks to dedicated processing resources to address one or more of the aforementioned problems and other potential issues. In this scheme, a Capsule Network model (CapsNN) is used to classify and train a training set. Capsules for each class are extracted from the trained CapsNN as distilled samples. Soft labels for the distilled samples are then generated using the original training set.

[0018] In this way, the distilled samples generated by CapsNN have high interpretability, and based on the nature of tag inheritance, the distilled samples can be easily mapped to their associated original samples, thus enabling them to be applied to applications such as image retrieval and data quality measurement.

[0019] Figure 1 A schematic diagram of an environment 100 in which embodiments of the present disclosure can be implemented is shown. (See diagram for reference.) Figure 1 As shown, environment 100 includes computing device 110, which may be a device with strong computing power, such as cloud server, smartphone, laptop, tablet, desktop computer or edge computing device.

[0020] like Figure 1 As shown, a capsule neural network model 130 is deployed in the computing device 110. Each capsule in the capsule neural network model 130 is composed of a set of vector neurons containing all relevant feature information within the capsule network structure. The length of the vector neuron's vector represents the probability of an entity's existence, and its direction represents the entity's instantiation parameters. The sum of the probabilities output by each capsule is not equal to 1, meaning the capsule has the ability to recognize multiple objects simultaneously and can segment highly overlapping digits. For example, the capsule neural network model 130 can be a machine learning model for classifying image samples. When the computing device 110 receives an image sample training set 120, it can use the capsule neural network model 130 to process the image sample training set 120 for classification training. After training, a classification rule 140 for classifying the samples can be obtained.

[0021] For example, in Figure 1In the example shown, the image sample training set 120 includes image samples of handwritten digits, specifically multiple handwritten Arabic numerals 0, 1, 2, 3, 4, 5, 6, 7, 8, and 9. The capsule neural network model 130 may include multiple layers. In the illustrated embodiment, the capsule neural network model 130 mainly consists of convolutional layers, primary capsule layers, and class capsule layers. The primary capsule layer is the first capsule layer, and the class capsule layer is the last capsule layer. Feature extraction from the image samples is performed by the convolutional layers, and the extracted features are then fed into the primary capsule layer. The activation function of the convolutional layers is the ReLU function. Feature transfer between capsule layers is accomplished through dynamic routing. Capsules can also be divided into different levels: lower-level capsules can be called primary capsules, and higher-level capsules can be called high-level capsules. Lower-level capsules extract pose parameters from pixels and create a part-whole hierarchical structure. Features from the primary capsules are continuously fused into higher-level capsules, ultimately resulting in a class capsule containing multiple features for a single category, and the number of class capsules is typically the same as the number of categories. For example, when training the capsule neural network model 130 using the image sample training set 120, which includes 10 categories, 10 category capsules will be generated in the capsule neural network model 130.

[0022] It should be understood that Figure 1 The image sample training set 120 and capsule neural network model 130 shown are merely examples, and the classification and number of samples, etc., are not to be construed as limiting the scope of the embodiments of this disclosure. Furthermore, Figure 1 The classification of the specific samples shown is merely exemplary and is not intended to be a limitation of this disclosure.

[0023] The following will combine Figures 2 to 4 The following describes in detail the exemplary embodiments of this disclosure. Figure 2 A flowchart of an example method 200 for generating image samples according to an embodiment of the present disclosure is shown. Figure 2 Method 200 in the middle, for example, can be derived from Figure 1 The computing device 110 in the middle performs the operation.

[0024] At point 202, computing device 110 processes a set of image samples using a classification capsule network model to obtain a set of distilled samples. The feature distribution of the set of distilled samples indicates the feature distribution of the set of image samples. In this embodiment, the classification capsule network model can be... Figure 1The capsule neural network model 130 is described above. As discussed, the number of classification capsules in the trained capsule network model is less than the number of samples, and due to the dynamic routing mechanism in the capsule network model, the classification capsules inherit the features of the training samples well and retain the correlation information between features. In some embodiments, a set of distilled samples can be obtained by extracting the classification capsules from the trained capsule network model, that is, the capsules in the last layer of the capsule layer. It should be understood that the classification capsules obtained by training with image samples are a set of vectors. These vectors can be visualized to obtain the corresponding images, thereby obtaining the distilled samples.

[0025] In some embodiments, when training a capsule network model, a corresponding loss equation can be set for training so that the obtained classification capsules have predetermined properties.

[0026] In some embodiments, the computing device 110 can process a set of image samples using a classification capsule network model based on the feature distribution of multiple features based on image samples and the feature distribution of a set of distilled samples to obtain a set of distilled samples. In this way, the training process of the classification capsule is adjusted based on the differences in the corresponding feature distributions, for example, to minimize the difference in feature distribution between the distilled samples and the image samples, so that the obtained distilled samples include as many features of the corresponding category as possible from the original samples, and so that the distilled samples can include as much meta-information as possible from the image samples.

[0027] In some embodiments, a mutual information term can be set in the loss function to ensure that the distilled samples include as much metadata as possible from the image samples:

[0028] (1)

[0029] in For the original image sample set, For the distillation sample set, For distilled samples, and For image samples. The feature distribution of the image sample set and . For distilled samples Based on image samples The probability of generation.

[0030] In some embodiments, the computing device 110 can process a set of image samples using a classification capsule network model based on the similarity between distilled samples, such that the number of identical features included in distilled samples of different categories is below a predetermined threshold. In such embodiments, in order for the classification capsules, i.e., the resulting distilled samples, to include relatively independent features, i.e., to include as few common features as possible between classification capsules, training can be performed by setting the following term in the loss function:

[0031] (2)

[0032] in This is a pre-defined similarity between distilled samples.

[0033] In some embodiments, training can be performed using the following loss function:

[0034] (3)

[0035] in Formula (2) is used. Formula (1) is used.

[0036] In this way, it is possible to ensure that the distilled samples include as much metadata as possible from the image samples, while also ensuring that the distilled samples include as few common features as possible.

[0037] At point 204, computing device 110 determines a soft label for each distillation sample in a set of distillation samples based on the labels of a set of image samples. A soft label represents the probability that a distillation sample belongs to each of multiple categories. To apply the extracted distillation samples to other tasks, labels need to be assigned to the distillation samples. For example, soft labels can be assigned to the original distillation samples based on the labels of the image samples, according to their relationship to the original image samples. The process of determining soft labels will be described below. Figure 3 Let me describe it in detail.

[0038] In this way, by proposing classification capsules as distillation samples from the trained capsule network model and assigning soft labels to the distillation samples based on the original image samples, it is possible to ensure that the extracted distillation samples have high interpretability. Furthermore, based on the genetic nature of the labels, it is easy to map the distillation samples to their associated original samples, thus enabling them to be applied to applications such as image retrieval and data quality measurement.

[0039] The following will refer to Figure 3 To describe in detail the process of determining the label.

[0040] Figure 3A flowchart of an example method 300 for determining a soft tag according to an embodiment of the present disclosure is shown. Figure 3 Method 300 in the middle, for example, can be derived from Figure 1 The computing device 110 in the middle performs the operation.

[0041] At 302, computing device 110 determines the correlation between each distillation sample in a set of distillation samples and each image sample in a set of image samples. The correlation may, for example, indicate the number or proportion of identical features included in the distillation sample and the image sample. The correlation may also indicate, for example, the similarity of the feature representations of the distillation sample and the image sample in the feature space. In some embodiments, the computing device 110 can obtain distillation features for each distillation sample while processing a set of image samples using a classification capsule network model. Distillation features may be, for example, feature vectors of the distillation sample in a specific feature space. The computing device 110 processes a set of image samples using a classification model to obtain the original features of each image sample in the set. Finally, the computing device 110 determines the correlation based on the distillation features and the original features. In this way, the correlation between the distillation features and the image samples can be quantitatively determined using feature vectors.

[0042] At position 304, the probability that each distillation sample belongs to each category is determined based on the correlation between each distillation sample and the image samples in each category.

[0043] In some embodiments, the computing device 110 may determine a weighted sum of the similarities between each distillation sample and image samples for each category, thereby using the weighted sum as the probability that each distillation sample belongs to each category. In some embodiments, normalization may also be performed after obtaining the probability.

[0044] At 306, computing device 110 can determine a soft label for each distillation sample based on the probability that each distillation sample belongs to each category.

[0045] In this way, by calculating the correlation between each distillation sample and all image samples in each category, the probability of a distillation sample belonging to the corresponding category can be represented by the correlation, thus obtaining a soft label. The following will refer to... Figure 4 Describe in detail the process of generating image samples.

[0046] Figure 4 A schematic diagram of an example process 400 for generating image samples according to some embodiments of the present disclosure is shown. Figure 4 For example, process 400 in the middle can be made by Figure 1 The computing device 110 in the middle performs the operation. For example... Figure 4 As shown, using Figure 1The image sample training set 120 in the image sample training set is used to classify and train the capsule neural network model 130. Figure 4 In the illustrated embodiment, the image sample training set 120 includes N image samples. After training, classification capsules are extracted from the trained capsule neural network model 130 to serve as the distillation sample set 410, and the distillation feature matrix 420 of the distillation sample set 410 is obtained simultaneously. The number of distilled samples in the distillation sample set 410 is d, and each distillation feature is an m-dimensional feature vector; therefore, the distillation feature matrix 420 is d×m dimensional. It should be understood that although in Figure 4 The illustrated embodiment includes 10 distillation samples corresponding to 10 categories, i.e., d = 10. However, for the sake of generality, m will continue to be used hereinafter to generally represent the number of distillation features. Then, the distillation feature matrix 420 is multiplied by an N×d correlation matrix 430 to obtain the pattern feature matrix 440 of the image sample training set 120, and the pattern feature matrix 440 is N×m. Each element in the correlation matrix 430 corresponds to the correlation between the distillation sample and the pattern sample.

[0047] To compute the relevance matrix 430, it is initialized, for example, by making it an N×d identity matrix. This yields the initialized pattern feature matrix 440. Next, a classification model is used to train the image sample training set 120, resulting in another N×m pattern feature matrix 440. The relevance matrix 430 is then derived from the pattern feature matrix 440 and the distillation feature matrix 420. The columns of the relevance matrix 430 are the relevance vectors 431. Each relevance vector 431 indicates the relevance of a distillation sample to each pattern sample. Finally, a weighted sum of the relevance vectors 431 is calculated based on the classification, resulting in a d-dimensional soft label 450.

[0048] In some alternative embodiments, an additional classification model can be used to train the image sample training set 120 and obtain the same m-dimensional pattern feature vector, thereby obtaining the pattern feature matrix 440 (represented by the dashed arrow), and then the correlation matrix 430 is obtained from the pattern feature matrix 440 and the distillation feature matrix 420.

[0049] Figure 5 A schematic block diagram of an example device 500 that can be used to implement embodiments of the present disclosure is shown. Figure 5As shown, device 500 includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 502 or loaded from storage unit 508 into random access memory (RAM) 503. RAM 503 may also store various programs and data required for the operation of device 500. CPU 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0050] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0051] The various processes and procedures described above, such as method 200, method 300, and process 400, can be executed by processing unit 501. For example, in some embodiments, method 200, method 300, and process 400 can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by CPU 501, one or more actions of method 200, method 300, and process 400 described above can be performed.

[0052] This disclosure can be a method, apparatus, system, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of this disclosure.

[0053] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0054] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0055] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0056] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0057] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0058] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0059] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0060] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for generating image samples, comprising: By processing a set of image samples using a classification capsule network model, a set of distilled samples is obtained, the feature distribution of which indicates the feature distribution of the set of image samples; and Based on the labels of the set of image samples, a soft label is determined for each distillation sample in the set of distillation samples, the soft label representing the probability that the distillation sample belongs to each of the multiple categories; The method for obtaining the set of distilled samples includes: The classification capsule network model is used to process the set of image samples based on a loss function associated with the similarity between the distillation samples of the set of distillation samples, such that the number of identical features included in distillation samples of different classifications is below a predetermined threshold.

2. The method of claim 1, wherein obtaining a set of distilled samples comprises: The set of distilled samples is obtained by using a classification capsule network model to process a set of image samples based on the feature distribution of multiple features of the set of image samples and the feature distribution of the set of distilled samples.

3. The method of claim 1, wherein determining the soft label for each distillation sample in the set of distillation samples comprises: Determine the correlation between each distillation sample in the set of distillation samples and each image sample in the set of image samples; Based on the correlation between each distillation sample and image samples in each category, determine the probability that each distillation sample belongs to each category; and A soft label is determined for each distillation sample based on the probability that each distillation sample belongs to each category.

4. The method of claim 3, wherein determining the relevance comprises: Obtain the distillation features of each distillation sample in the set of distillation samples, wherein the distillation features are obtained when processing a set of image samples using a classification capsule network model; and By processing the set of image samples using a classification model, the original features of each image sample in the set of image samples are obtained; and The correlation is determined based on the distillation characteristics and the original characteristics.

5. The method of claim 3, wherein determining the probability that each distilled sample belongs to each category comprises: The weighted sum of the similarities between each distillation sample and the image samples in each category is determined as the probability that each distillation sample belongs to each category.

6. The method according to claim 1, further comprising: Use distilled samples and soft labels to train machine learning models.

7. An electronic device, comprising: processor; as well as A memory coupled to the processor, the memory having instructions stored therein, the instructions causing the device to perform actions when executed by the processor, the actions including: By processing a set of image samples using a classification capsule network model, a set of distilled samples is obtained, the feature distribution of which indicates the feature distribution of the set of image samples; and Based on the labels of the set of image samples, a soft label is determined for each distillation sample in the set of distillation samples, the soft label representing the probability that the distillation sample belongs to each of the multiple categories; The method for obtaining the set of distilled samples includes: The classification capsule network model is used to process the set of image samples based on a loss function associated with the similarity between the distillation samples of the set of distillation samples, such that the number of identical features included in distillation samples of different classifications is below a predetermined threshold.

8. The electronic device of claim 7, wherein obtaining a set of distilled samples comprises: The set of distilled samples is obtained by using a classification capsule network model to process a set of image samples based on the feature distribution of multiple features of the set of image samples and the feature distribution of the set of distilled samples.

9. The electronic device of claim 7, wherein determining the soft tag for each distillation sample in the set of distillation samples comprises: Determine the correlation between each distillation sample in the set of distillation samples and each image sample in the set of image samples; Based on the correlation between each distillation sample and image samples in each category, determine the probability that each distillation sample belongs to each category; and A soft label is determined for each distillation sample based on the probability that each distillation sample belongs to each category.

10. The electronic device of claim 9, wherein determining the relevance comprises: Obtain the distillation features of each distillation sample in the set of distillation samples, wherein the distillation features are obtained when processing a set of image samples using a classification capsule network model; and By processing the set of image samples using a classification model, the original features of each image sample in the set of image samples are obtained; and The correlation is determined based on the distillation characteristics and the original characteristics.

11. The electronic device of claim 9, wherein determining the probability that each distilled sample belongs to each category comprises: The weighted sum of the similarities between each distillation sample and the image samples in each category is determined as the probability that each distillation sample belongs to each category.

12. The electronic device of claim 7, wherein the action further comprises: Use distilled samples and soft labels to train machine learning models.

13. A computer program product tangibly stored on a computer-readable medium and comprising machine-executable instructions that, when executed, cause a machine to perform the method according to any one of claims 1 to 6.

Images