Information processing apparatus and information processing method

Abstract

The present disclosure relates to an information processing apparatus and an information processing method. The information processing apparatus is used to train a classification model capable of classifying second samples without labels of a second domain different from a first domain based on first samples with labels of the first domain, there being a same shared class between classes of the second samples and classes of the first samples. The information processing apparatus includes: a clustering unit that clusters each of the second samples toward its adjacent first sample or second sample; a plurality of first classification units that each classifies the clustered second sample based on a partial class of the first samples; and an identification unit that identifies the second sample without the shared class based on classification results of the plurality of first classification units, wherein the classification model is trained using a clustering result of the clustering unit and an identification result of the identification unit. According to the information processing technology of the present disclosure, the learning effect of transfer learning can be significantly improved.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of information processing technology. In particular, embodiments of this disclosure relate to information processing apparatus and methods for transfer learning between different domains, and classification models trained using said information processing methods. Background Technology

[0002] In recent years, deep learning has made great progress in many machine learning tasks and applications. With the emergence of more and more machine learning application scenarios, although supervised machine learning has shown good performance, training requires a large number of labeled samples. However, labeling samples is a time-consuming and laborious task. Therefore, transfer learning, which can apply knowledge learned in one domain (hereinafter referred to as the source domain) to another different but related domain (hereinafter referred to as the target domain), has received increasing attention.

[0003] Furthermore, many cross-domain machine learning tasks are based on the assumption that the training sample set (i.e., the set of labeled samples in the source domain) and the test set (i.e., the set of unlabeled samples in the target domain) share the same class. However, since there are often differences between the classes of samples in the target domain and the classes of samples in the source domain in real-world applications—that is, there are inter-domain class gaps—this assumption does not hold true in many application scenarios.

[0004] Figures 1A to 1D This diagram illustrates various cases regarding the class gap between the source and target domains. Figures 1A to 1D In this context, the source domain can be the domain of a photographic image, while the target domain can be the domain of a cartoon image. Furthermore, as... Figures 1A to 1D As shown, for example, images in the source and target domains can have four labels, or four categories: Bed, Chair, Bike, and Fork. Specifically, Figure 1A This illustrates an application scenario known as Closed-set Domain Adaptation (CDA), where the categories of the target domain are exactly the same as those of the source domain; that is, the categories of the target domain are the complete set of the categories of the source domain, and therefore there are no category gaps. Figure 1B This illustrates an application scenario of Partial Domain Adaptation (PDA), where the categories of the target domain are a subset of the categories of the source domain. Therefore, there are samples of unknown categories in the target domain, meaning there is a category gap between the source and target domains. Figure 1CThis illustrates an application scenario for Open-set Domain Adaptation (OSDA), where the category of the source domain is a subset of the category of the target domain, meaning the category of the source domain is a subset of the category of the target domain, thus creating a category gap between the source and target domains. Figure 1D This illustrates an application scenario for Open Partial Domain Adaptation (OPDA), where some categories of the target domain are the same as some categories of the source domain; that is, there is an overlap between the categories of the source and target domains, thus creating a category gap between the source and target domains. Figures 1A to 1D As shown, for applications suitable for transfer learning, even in the presence of class gaps, there should be at least an intersection between the classes in the source domain and the classes in the target domain, i.e., the same shared classes.

[0005] Therefore, in the existing technology, there is still a need to provide a generalized information processing technique for inter-domain transfer learning, which can address, for example... Figures 1A to 1D The various application scenarios shown demonstrate good transfer learning performance. Summary of the Invention

[0006] A brief overview of this disclosure is provided below to offer a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the disclosure. It is not intended to identify key or essential parts of the disclosure, nor is it intended to limit its scope. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0007] To address the problems in the existing technologies mentioned above, it is necessary to provide a generalized information processing technology applicable to various application scenarios that enables inter-domain transfer learning, for the accurate and rapid detection of unlabeled samples.

[0008] To achieve the objectives of this disclosure, according to one aspect of this disclosure, an information processing apparatus is provided for training a classification model capable of classifying unlabeled second samples in a second domain, different from the first domain, based on labeled first samples in a first domain, wherein there is a shared category between the category of the second samples and the category of the first samples. The information processing apparatus includes: a clustering unit configured to cluster each of the second samples toward its adjacent first or second sample; a plurality of first classification units, each first classification unit configured to classify the clustered second samples based on a partial category of the first samples; and an identification unit configured to identify second samples that do not share a category based on the classification results of the plurality of first classification units, wherein the classification model is trained using the clustering results of the clustering units and the identification results of the identification units.

[0009] According to another aspect of this disclosure, an information processing method is provided for training a classification model capable of classifying unlabeled second samples in a second domain, different from the first domain, based on labeled first samples in a first domain, wherein there is a shared category between the categories of the second samples and the categories of the first samples. The information processing method includes: clustering each of the second samples toward its neighboring first or second samples; classifying the clustered second samples based on a partial category of the first samples; identifying second samples that do not share a category based on the classification results; and training the classification model using the clustering results and the identification results.

[0010] According to another aspect of this disclosure, a classification model trained using the information processing method described above is also provided.

[0011] According to another aspect of this disclosure, a computer program capable of implementing the above-described information processing method is also provided.

[0012] In addition, a computer program product having at least a computer-readable storage medium is provided, on which computer program code for implementing the above-described information processing method is recorded.

[0013] This disclosure proposes a novel information processing technique for training classification models that implement transfer learning. Classification models trained using this technique can achieve good transfer learning performance in various application scenarios (with or without inter-domain class gaps), and can significantly improve detection accuracy compared to traditional classification models that implement transfer learning. Attached Figure Description

[0014] The above and other objects, features, and advantages of this disclosure will be more readily understood by referring to the following description of embodiments of the present disclosure in conjunction with the accompanying drawings, in which:

[0015] Figures 1A to 1D Schematic diagrams illustrating various cases of class gaps between the source and target domains.

[0016] Figure 2 A block diagram of an information processing apparatus according to an embodiment of the present disclosure is shown;

[0017] Figure 3 A schematic diagram of the structure of an information processing apparatus according to an embodiment of the present disclosure is shown;

[0018] Figure 4 A schematic diagram illustrating the training effect of an information processing apparatus according to an embodiment of the present disclosure is shown;

[0019] Figure 5 A schematic diagram illustrating the training effect of an information processing apparatus according to an embodiment of the present disclosure is shown;

[0020] Figure 6 A flowchart illustrating an information processing method according to an embodiment of this disclosure is shown; and

[0021] Figure 7 A simplified structural diagram of a general-purpose machine for implementing the information processing method according to embodiments of the present disclosure is shown. Detailed Implementation

[0022] In the following description, some embodiments of this disclosure will be described in detail with reference to the accompanying illustrative drawings. When elements are indicated by reference numerals in the drawings, the same elements will be represented by the same reference numerals even if the same elements are shown in different drawings. Furthermore, in the following description of this disclosure, detailed descriptions of known functions and configurations incorporated herein will be omitted where such omissions may obscure the subject matter of this disclosure.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context otherwise indicates. It will also be understood that the terms “comprising,” “including,” and “having” as used in the specification are intended to specifically describe the presence of the stated features, entities, operations, and / or components, but do not preclude the presence or addition of one or more other features, entities, operations, and / or components.

[0024] Unless otherwise defined, all terms used herein, including technical and scientific terms, shall have the same meaning as commonly understood by one of ordinary skill in the art to which the inventive concept of this disclosure pertains. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant field, and should not be interpreted in an idealized or overly formal sense unless explicitly defined herein.

[0025] In the following description, numerous specific details are set forth to provide a full understanding of this disclosure. This disclosure may be practiced without some or all of these specific details. In other examples, to avoid obscuring the disclosure with unnecessary detail, only components closely related to the embodiments according to this disclosure are shown in the accompanying drawings, while other details less relevant to this disclosure are omitted.

[0026] The problem addressed in this disclosure concerns cross-domain adaptation in transfer learning, also known as Universal Domain Adaptation (UniDA). In this paper, the source domain (hereinafter referred to as the first domain) contains labeled samples (hereinafter referred to as first samples), while the target domain (hereinafter referred to as the second domain) contains unlabeled samples (hereinafter referred to as second samples). Furthermore, as described above, in applications suitable for transfer learning, the categories of the second samples and the first samples share a common category; that is, at least a portion of the categories of the second samples in the target domain are the same as at least a portion of the categories of the first samples in the source domain. Moreover, in this paper, the second samples in the target domain with shared categories are referred to as known-category second samples, while the second samples in the target domain without shared categories are referred to as unknown-category second samples.

[0027] To address the problems existing in the prior art described above, this paper proposes an information processing technique for training classification models that implement transfer learning. The classification model trained by this technique can achieve good transfer learning performance for various application scenarios (with or without inter-domain class gaps), and can significantly improve detection accuracy compared with traditional classification models that implement transfer learning.

[0028] The following is combined Figures 2 to 5 The information processing apparatus 200 according to embodiments of the present disclosure will be described in detail.

[0029] Figure 2 A block diagram of an information processing apparatus 200 according to an embodiment of the present disclosure is shown. Figure 3A schematic diagram showing the structure of an information processing apparatus 200 according to an embodiment of the present disclosure is provided. Figure 4 and Figure 5 A schematic diagram illustrating the principle of an information processing apparatus 200 according to an embodiment of the present disclosure is shown.

[0030] According to embodiments of the present disclosure, the information processing apparatus 200 is used to train a classification model 210. The classification model 210 is capable of classifying unlabeled second samples in a second domain, different from the first domain, based on labeled first samples in a first domain, where there is a shared category between the categories of the second samples and the first samples. The information processing apparatus 200 may include: a clustering unit 201 configured to cluster each of the second samples toward its neighboring first or second sample; a plurality of first classification units 202, each first classification unit 202 configured to classify the clustered second samples based on a partial category of the first samples; and an identification unit 203 configured to identify second samples that do not share a category based on the classification results of the plurality of first classification units. According to embodiments of the present disclosure, the classification model 210 can be trained using the identification results of the identification unit 203.

[0031] According to embodiments of this disclosure, the first sample and the second sample can be images. That is, the sample in the source domain (the first sample) is a labeled image, while the sample in the target domain (the second sample) is an unlabeled image. Those skilled in the art will recognize that although exemplary embodiments of this disclosure are described herein using images as samples, the scope of this disclosure is not limited thereto. Based on the teachings of this disclosure, those skilled in the art can envision that the information processing apparatus according to this disclosure can also be applied to other machine learning application areas besides image processing, such as speech analysis, natural language processing, and weather forecasting.

[0032] As described above, in the UniPDA problem, the specific category of the second sample is unknown. The classification model 210 needs to identify whether the second sample is a known category second sample or an unknown category second sample, and, if the second sample is a known category second sample, to identify its specific category. The information processing apparatus 200 according to embodiments of this disclosure can train the classification model 210 to achieve the above functions.

[0033] According to embodiments of this disclosure, the classification model 210 can be implemented using a convolutional neural network (CNN) model. Given that convolutional neural network (CNN) models are known to those skilled in the art, their details are not described in more detail herein for the sake of brevity.

[0034] Specifically, according to embodiments of this disclosure, the classification model 210 may include a feature extraction unit 211 and a second classification unit 212. The feature extraction unit 211 may extract feature values ​​of the first sample and the second sample. The second classification unit 212 may classify the first sample and the second sample based on all categories of the first sample.

[0035] As an example, not a limitation, such as Figure 3 As shown, the feature extraction unit 211 can be implemented using a ResNet-50 model pre-trained on the ImageNet image database as the backbone. For example, as Figure 3 As shown, the input to the feature extraction unit 211 can be a 256×256 pixel image. The feature extraction unit 211 can extract features representing the image, which can be represented by a 2048-dimensional vector. Specifically, according to embodiments of this disclosure, the feature extraction unit 211 can normalize the second sample to facilitate subsequent processing. Since normalization is known to those skilled in the art, its details are not described in more detail herein for the sake of brevity.

[0036] Since the structure of the classification model 210 (including feature extraction unit 211 and second classification unit 212) implemented based on the CNN model is known to those skilled in the art, for the sake of brevity, its details will not be described in more detail here.

[0037] According to embodiments of this disclosure, the clustering unit 201 can perform clustering based on the feature quantities of the first and second samples extracted by the feature extraction unit 211. According to this disclosure, as... Figure 4 As shown, the clustering performed by clustering unit 201 aims to cluster samples (second samples) in the target domain toward their neighboring samples. It should be noted that in the feature space, the neighboring samples of a sample in the target domain can be samples in the source domain (first samples) or other samples in the target domain (other second samples).

[0038] Specifically, Figure 4 (a) shows a schematic diagram of the distribution of the first and second samples in the feature space before clustering, while Figure 4 (b) shows a schematic diagram of the distribution of the first and second samples in the feature space after clustering. Figure 4 In the diagram, triangles represent samples in the source domain (first samples), while circles represent samples in the target domain (second samples). Furthermore, hollow triangles represent first samples with a shared class, while hollow circles represent second samples with a shared class. Correspondingly, solid triangles represent first samples without a shared class, while solid circles represent second samples without a shared class.

[0039] like Figure 4 As shown in (b), the clustering process of clustering unit 201 can cause each of the second samples to cluster toward its neighboring first or second sample to improve the discrimination.

[0040] According to embodiments of this disclosure, for each of the second samples, the clustering unit may cluster the sample based on the distance between that sample and its neighboring first or second samples. Specifically, according to embodiments of this disclosure, for each of the second samples, the clustering unit may determine the neighboring samples of that sample by comparing the distance between that sample and all the first and second samples.

[0041] For example, suppose It is the feature quantity of the source domain sample (first sample) extracted by feature extraction unit 211, where d is the dimension of the feature vector, m s This is the number of samples in the source domain. Accordingly, we assume... It is the feature quantity of the target domain sample (second sample) extracted by feature extraction unit 211, where m t This is the number of samples in the target domain. As mentioned above, the feature quantity V s and V t It can be normalized. Feature V s and V t They can be represented together as Features of the sample in the i-th normalized target domain The j-th feature V j and The probability of adjacent pairs can be determined by γ. i，j It means that V j ∈V, j≠i. According to the implementation of this public content, probability γ i，j It can be limited by the following formula (1):

[0042]

[0043] In equation (1), φ represents the distance between the eigenvectors in the feature space. According to embodiments of this disclosure, this distance can be a cosine distance or an Euclidean distance.

[0044] For example, clustering loss It can be represented by the following formula (2).

[0045]

[0046] Subsequently, each of the plurality of first classification units 202 can classify the clustered second samples based on a subset of the categories of the first samples. It should be noted that the difference between the first classification unit 202 and the second classification unit 212 is that the first classification unit 202 classifies only a subset of the categories of the samples in the source domain, while the second classification unit 212 classifies all the categories of the samples in the source domain.

[0047] According to embodiments of this disclosure, the number of partial categories in the plurality of first classification units 202 is the same for each other. According to embodiments of this disclosure, the number of the plurality of first classification units 202 is equal to the total number of categories in the first sample divided by the number of partial categories in each first classification unit 202. Furthermore, according to embodiments of this disclosure, the set of partial categories in the plurality of first classification units 202 is the complete set of all categories in the first sample.

[0048] For example, assuming J is the number of multiple first classification units 202, the set of multiple first classification units 202 can be represented as {h1, h2, ..., h...} J Furthermore, assume K is the number of categories of the samples in the source domain, i.e., the number of labels of the first sample. Thus, for each first classification unit 202, K / J categories can be discarded sequentially. Furthermore, assume h represents a second classification unit 212 that classifies samples based on all categories of the samples in the source domain. For example, suppose the first sample has 6 categories 1 to 6, and there are 3 first classification units 202. Thus, the first first classification unit 202 can discard categories 1 and 2, classifying only based on categories 3 to 6. Furthermore, the second first classification unit 202 can discard categories 3 and 4, classifying only based on categories 1 and 2, and categories 5 and 6. Furthermore, the third first classification unit 202 can discard categories 5 and 6, classifying only based on categories 1 to 4.

[0049] According to embodiments of this disclosure, the identification unit 203 can identify a second sample that does not share a category based on the difference between the classification results of multiple first classification units 202 and the classification results of the second classification unit 212.

[0050] According to the embodiments of this disclosure, the identification unit 203 can count the number of first classification units 202 whose classification results are different from those of the second classification unit 212 for each sample in the second sample, and determine that the second sample does not have a shared category when the counted number is greater than a preset threshold.

[0051] For example, an indicator function I(h(g(x)),h can be defined. j(g(x))), where x represents the second sample, g(x) represents the clustered feature of the second sample, and h(g(x)) represents the classification result of the second classification unit 212. j (g(x)) represents the classification result of the j-th first classification unit 202, where 1≤j≤J. For each second sample x, the indicator function I is defined in h(g(x)) and h... j (g(x)) takes the value 1 if they are not both 1, otherwise it takes the value 0.

[0052] For the second sample x i , where 1≤x i ≤m t You can use ω i This indicates that the second sample x i Is it a second sample that does not share a class, i.e., a second sample of an unknown class? Specifically, ω i It can be limited by the following equation (3). As shown in equation (3), if ω i If the value is 1, then the corresponding second sample x can be considered as... i It is a second sample that does not share a category.

[0053]

[0054] Wherein β is a preset threshold, and its value ranges from 1 / J to 1. According to the embodiments of this disclosure, the preset threshold β can be determined according to the application scenario. In this respect, the larger the preset threshold β, the stricter the constraint on detecting the second sample as an unknown category second sample. For example, in the field of finance, it is desirable to have the highest possible detection accuracy for second samples that do not share a category in order to avoid financial losses, so the preset threshold β can be set to a larger value.

[0055] According to the embodiments of this disclosure, the identification unit 203 can calculate the identification loss as shown in the following formula (4).

[0056]

[0057] Where ρ i，k Indicates the second sample x i The probability of being predicted as class k by the second classification unit h is output by the second classification unit h. This clustering loss is introduced during the training of the classification model 210. and recognition loss This can improve the ability of classification model 210 to identify second samples that do not share a class. In this respect, the purpose of using multiple first classification units 201 is to identify second samples that do not share a class.

[0058] Specifically, Figure 5Figure (a) shows the entropy-based distribution of the second classification unit 212 based on all categories of the first sample. As is known to those skilled in the art, the higher the certainty of the classification result, the smaller the entropy, and vice versa. For example, entropy can be represented by the following equation (5).

[0059]

[0060] Ideally, the second sample, which does not share a class, should be located at the position of maximum entropy, meaning it does not belong to any class of the first sample. However, due to factors such as initial value settings, etc., Figure 5 As shown in (a), some second samples that do not share a category may not be located at the position of maximum entropy. According to this disclosure, by introducing multiple first classification units 202 to remove some categories from some first samples to introduce new constraints, the entropy distribution of second samples that do not share a category can be adjusted, thereby enabling more accurate detection of second samples that do not share a category. For example, as... Figure 5 As shown in (b), by removing the category of the first sample in the lower left corner, the second sample in the upper right corner, which does not share a category, can be placed at the position of maximum entropy, thereby improving classification accuracy.

[0061] The classification model trained using the information processing technology disclosed herein can achieve good transfer learning performance for various application scenarios (with or without inter-domain class gaps), and can significantly improve detection accuracy compared to traditional classification models that implement transfer learning.

[0062] For example, the experiment is based on the Office-Home image database commonly used in existing technologies, and the classification model 210 is trained using the information processing device 200 according to the present invention.

[0063] Experimental results show that the classification model 210 trained by the information processing device 200 according to the present invention can achieve an average detection accuracy of up to 82.7% for samples in the target domain (when trained using three second classification units 202), which is at least 2.4% higher than the methods of the prior art.

[0064] Accordingly, this disclosure also proposes an information processing method for training a classification model that implements transfer learning. This classification model is capable of classifying unlabeled second samples in a second domain (different from the first domain) based on a labeled first sample in a first domain, where the categories of the second samples and the first samples share the same category.

[0065] Figure 6 This is a flowchart illustrating an information processing method 600 according to an embodiment of the present disclosure.

[0066] Information processing method 600 begins at step S601. Subsequently, in step S602, each of the second samples is clustered toward its neighboring first or second sample. According to embodiments of this disclosure, the processing in step S602 can be performed, for example, by referring to the above description. Figures 2 to 5 The clustering unit 201 described herein will be used to implement this, and its details will not be elaborated here.

[0067] Subsequently, in step S603, the clustered second samples are classified based on a subset of the categories of the first samples. According to embodiments of this disclosure, the processing in step S603 can be performed, for example, by referring to the above description. Figures 2 to 5 The first classification unit 202 described is used to implement this, so its details will not be elaborated here.

[0068] Subsequently, in step S604, a second sample that does not share a category is identified based on the classification result. According to embodiments of this disclosure, the processing in step S604 can, for example, be performed according to the above reference... Figures 2 to 5 The identification unit 203 described herein will not be elaborated upon here.

[0069] Subsequently, in step S605, the clustering results and identification results are used to train a classification model. According to embodiments of this disclosure, the processing in step S604 can, for example, be performed according to the above reference... Figures 2 to 5 The clustering unit 201, the first classification unit 202, and the identification unit 203 described are implemented through the interaction with the classification model 210.

[0070] Finally, the information processing method 600 ends at step S606.

[0071] The information processing techniques disclosed herein can significantly improve the learning outcomes of transfer learning.

[0072] Figure 7 This is a simplified structural diagram illustrating a general-purpose machine 700 that can be used to implement the information processing method and information processing apparatus according to embodiments of the present disclosure. The general-purpose machine 700 may be, for example, a computer system. It should be noted that the general-purpose machine 700 is merely an example and does not imply any limitation on the scope or functionality of the information processing method and information processing apparatus of the present disclosure. Nor should the general-purpose machine 700 be construed as having a dependency on or requirement for any component or combination thereof shown in the above-described information processing method or information processing apparatus.

[0073] exist Figure 7In this system, the Central Processing Unit (CPU) 701 performs various processes based on programs stored in the Read-Only Memory (ROM) 702 or programs loaded into the Random Access Memory (RAM) 703 from the storage section 708. The RAM 703 also stores images required as needed when the CPU 701 performs various processes, etc. The CPU 701, ROM 702, and RAM 703 are connected to each other via a bus 704. An input / output interface 705 is also connected to the bus 704.

[0074] The following components are also connected to the input / output interface 705: input section 706 (including keyboard, mouse, etc.), output section 707 (including display, such as cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.), storage section 708 (including hard disk, etc.), and communication section 709 (including network interface card, such as LAN card, modem, etc.). The communication section 709 performs communication processing via a network, such as the Internet. Drive 710 may also be connected to the input / output interface 705 as needed. Removable media 711, such as disk, optical disk, magneto-optical disk, semiconductor memory, etc., can be installed on drive 710 as needed, so that computer programs read from them can be installed into storage section 708 as required.

[0075] When the above series of processes are implemented by software, the program constituting the software can be installed from a network such as the Internet or from a storage medium such as removable medium 711.

[0076] Those skilled in the art will understand that such storage media are not limited to Figure 7 The illustration shows a removable medium 711 containing a program, distributed separately from the device to provide the program to the user. Examples of removable media 711 include magnetic disks (including floppy disks), optical disks (including optical disc read-only memory (CD-ROM) and digital versatile disks (DVDs)), magneto-optical disks (including mini-discs (MDs) (registered trademarks)), and semiconductor memory. Alternatively, the storage medium may be a ROM 702, a hard disk included in storage section 708, etc., containing programs and distributed to the user along with the device containing them.

[0077] Furthermore, this disclosure also proposes a program product storing machine-readable instruction code. When the instruction code is read and executed by a machine, the information processing method described above according to this disclosure can be performed. Accordingly, the various storage media listed above for carrying such a program product are also included within the scope of this disclosure.

[0078] The foregoing has been described in detail with reference to block diagrams, flowcharts, and / or embodiments, illustrating specific implementations of the apparatus and / or methods according to the present disclosure. When these block diagrams, flowcharts, and / or embodiments contain one or more functions and / or operations, those skilled in the art will understand that the various functions and / or operations in these block diagrams, flowcharts, and / or embodiments can be implemented individually and / or jointly by various hardware, software, firmware, or substantially any combination thereof. In one embodiment, several portions of the subject matter described herein may be implemented using application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), or other integrated forms. However, those skilled in the art will recognize that some aspects of the embodiments described in this specification can be implemented, in whole or in part, as one or more computer programs running on one or more computers (e.g., as one or more computer programs running on one or more computer systems), as one or more programs running on one or more processors (e.g., as one or more programs running on one or more microprocessors), as firmware, or as substantially any combination thereof, and that designing circuitry for this disclosure and / or writing software and / or firmware for this disclosure, based on the disclosure herein, is entirely within the capabilities of those skilled in the art.

[0079] In summary, the following technical solutions are provided in the embodiments according to this disclosure, but are not limited to these technical solutions:

[0080] Solution 1. An information processing apparatus for training a classification model, the classification model being capable of classifying unlabeled second samples in a second domain, different from the first domain, based on a first sample with a label in a first domain, wherein there is a shared category between the category of the second sample and the category of the first sample, the information processing apparatus comprising:

[0081] The clustering unit is configured to cause each of the second samples to cluster toward its adjacent first or second sample;

[0082] Multiple first classification units, each configured to classify clustered second samples based on a subset of the categories of the first samples; and

[0083] The identification unit is configured to identify a second sample that does not have the shared category based on the classification results of the plurality of first classification units.

[0084] The classification model is trained using the clustering results of the clustering units and the recognition results of the recognition units.

[0085] Option 2. The information processing device according to Option 1, wherein,

[0086] The classification model is implemented using a convolutional neural network model.

[0087] Option 3. The information processing device according to Option 2, wherein,

[0088] The classification model includes a feature extraction unit and a second classification unit.

[0089] The feature extraction unit is configured to extract feature quantities from the first sample and the second sample, and

[0090] The second classification unit is configured to classify the first sample and the second sample based on all categories of the first sample.

[0091] Option 4. The information processing device according to Option 3, wherein,

[0092] The feature extraction unit is configured to normalize the first sample and the second sample.

[0093] Option 5. The information processing device according to Option 3, wherein,

[0094] The clustering unit is configured to perform clustering based on the feature values ​​of the first sample and the second sample extracted by the feature extraction unit.

[0095] Option 6. The information processing apparatus according to Option 1, wherein,

[0096] For each of the second samples, the clustering unit is configured to cluster based on the distance between that sample and its neighboring samples.

[0097] Option 7. The information processing apparatus according to Option 6, wherein,

[0098] The distance mentioned is a cosine distance or an Euclidean distance.

[0099] Option 8. The information processing apparatus according to Option 1, wherein,

[0100] For each of the second samples, the clustering unit is configured to determine the neighboring samples of that sample by comparing the distance between that sample and all the first and second samples.

[0101] Option 9. The information processing apparatus according to Option 1, wherein,

[0102] The number of partial categories in the plurality of first classification units is the same for each other.

[0103] Option 10. The information processing apparatus according to Option 9, wherein,

[0104] The number of the plurality of first classification units is equal to the total number of categories of the first sample divided by the number of partial categories of each first classification unit.

[0105] Option 11. The information processing apparatus according to Option 1, wherein,

[0106] The set of partial categories of the multiple first classification units is the complete set of all categories of the first sample.

[0107] Option 12. The information processing apparatus according to Option 3, wherein,

[0108] The identification unit is configured to identify a second sample that does not have the shared category based on the difference between the classification results of the plurality of first classification units and the classification results of the second classification unit.

[0109] Option 13. The information processing apparatus according to Option 12, wherein the identification unit is configured to:

[0110] For each sample in the second sample, the number of first classification units whose classification results are different from those of the second classification unit is counted, and when the counted number is greater than a preset threshold, it is determined that the second sample does not have the shared category.

[0111] Option 14. The information processing device according to Option 13, wherein the preset threshold is determined according to the application scenario.

[0112] Scheme 15. The information processing apparatus according to any one of Schemes 1 to 14, wherein the first sample and the second sample are images.

[0113] Solution 16. An information processing method for training a classification model, said classification model being able to classify unlabeled second samples in a second domain, different from the first domain, based on a first sample with a label in a first domain, wherein there is a shared category between the category of the second sample and the category of the first sample, said information processing method comprising:

[0114] Each of the second samples clusters toward its adjacent first or second sample;

[0115] The clustered second sample is classified based on a subset of the categories of the first sample;

[0116] Identify a second sample that does not belong to the shared category based on the classification results; and

[0117] The classification model is trained using the clustering results and the identification results.

[0118] Scheme 17. A classification model for classifying a second sample without a label in a second domain, which is different from the first domain, based on a first sample with a label in a first domain, wherein there is a shared category between the category of the second sample and the category of the first sample, said classification model being trained by the information processing method according to claim 16.

[0119] Although this disclosure has been described above through a specific embodiment, it should be understood that those skilled in the art can design various modifications, improvements, or equivalents to the embodiments of this disclosure within the spirit and scope of the appended claims. These modifications, improvements, or equivalents should also be considered to be included within the protection scope of this disclosure.

Claims

1. An information processing apparatus for training a classification model, the classification model being capable of classifying unlabeled second samples in a second domain, different from the first domain, based on a first sample with a label in a first domain, wherein the categories of the second samples and the categories of the first samples share the same category, the information processing apparatus comprising: The clustering unit is configured to cause each of the second samples to cluster toward its adjacent first or second sample; Multiple first classification units, each first classification unit being configured to classify clustered second samples based on a subset of the categories of the first samples; as well as The identification unit is configured to identify a second sample that does not have the shared category based on the classification results of the plurality of first classification units. The classification model is trained using the clustering results of the clustering units and the recognition results of the recognition units. The first sample and the second sample are images.

2. The information processing apparatus according to claim 1, wherein, The classification model is implemented using a convolutional neural network model.

3. The information processing apparatus according to claim 2, wherein, The classification model includes a feature extraction unit and a second classification unit. The feature extraction unit is configured to extract feature quantities from the first sample and the second sample, and The second classification unit is configured to classify the first sample and the second sample based on all categories of the first sample.

4. The information processing apparatus according to claim 1, wherein, For each sample in the second sample, the clustering unit is configured to cluster based on the distance between that sample and its neighboring samples.

5. The information processing apparatus according to claim 1, wherein, The number of partial categories in the plurality of first classification units is the same for each other.

6. The information processing apparatus according to claim 5, wherein, The number of the plurality of first classification units is equal to the total number of categories of the first sample divided by the number of partial categories of each first classification unit.

7. The information processing apparatus according to claim 3, wherein, The identification unit is configured to identify a second sample that does not have the shared category based on the difference between the classification results of the plurality of first classification units and the classification results of the second classification unit.

8. The information processing apparatus according to claim 7, wherein, The identification unit is configured to: for each sample in the second sample, count the number of first classification units whose classification results are different from those of the second classification unit, and determine that the second sample does not have the shared category when the counted number is greater than a preset threshold.

9. An information processing method for training a classification model, said classification model being able to classify unlabeled second samples in a second domain, different from the first domain, based on a first sample with a label in a first domain, wherein there is a shared category between the category of the second sample and the category of the first sample, said information processing method comprising: Each of the second samples clusters toward its adjacent first or second sample; The clustered second sample is classified based on a subset of the categories of the first sample; Identify a second sample that does not belong to the shared category based on the classification results; as well as The classification model is trained using the clustering results and the identification results. The first sample and the second sample are images.

Images