Information processing device and information processing method

Abstract

The present disclosure relates to an information processing device and an information processing method. The information processing device is used to classify a second sample without a label in a second domain different from the first domain based on a first sample with a label in the first domain, and the category of the second sample is a subset or a full set of the category of the first sample. The information processing device includes: a first judgment unit for judging whether the category of the first sample is a shared category between the first domain and the second domain; and a second judgment unit for judging the shared category to which the second sample belongs, wherein the first judgment unit and the second judgment unit perform joint training by exchanging information. 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 technical field of information processing, and in particular, to an information processing apparatus and method for transfer learning between different domains, and a computer-readable storage medium. Background Art

[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 labor-intensive 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] In addition, many machine learning tasks are based on the assumption that the training sample set (i.e., the set of labeled samples) and the test set (i.e., the set of unlabeled samples) have the same sample distribution. However, in practical applications, this assumption may not hold true, especially in transfer learning. In transfer learning, samples collected from different domains will have a strong domain shift. There are many factors that lead to domain shift, such as changes in the environment, changes in sampling methods, and different data patterns of samples. In addition, the different sample label distributions (i.e., the category distribution of samples) between the source domain and the target domain will also affect the training of the model. All of the above factors will affect the accuracy of the final transfer learning model. Summary of the Invention

[0004] A brief overview of the present disclosure will be provided below to provide a basic understanding of certain aspects of the present disclosure. It should be understood that this overview is not an exhaustive overview of the present disclosure. It is not intended to identify key or important aspects of the present disclosure, nor is it intended to limit the scope of the present disclosure. Its purpose is simply to present certain concepts in a simplified form as a prelude to a more detailed description that will be discussed later.

[0005] In order to solve the problems in the prior art mentioned above, it is necessary to provide an information processing technology for transfer learning between different domains, so as to accurately and quickly realize the detection of unlabeled samples.

[0006] In order to achieve the purpose of the present disclosure, according to one aspect of the present disclosure, an information processing device is provided for classifying a second sample without a label in a second domain different from the first domain based on a first sample with a label in the first domain, wherein the category of the second sample is a subset or a full set of the category of the first sample, the information processing device comprising: a first judgment unit, configured to judge whether the category of the first sample is a shared category between the first domain and the second domain; and a second judgment unit, configured to judge the shared category to which the second sample belongs, wherein the first judgment unit and the second judgment unit perform joint training by exchanging information.

[0007] According to another aspect of the present disclosure, there is provided an information processing method for classifying a second sample without a label from a second domain different from the first domain based on a first sample with a label from the first domain, wherein the category of the second sample is a subset or a full set of the category of the first sample, the information processing method comprising: determining, by a first judgment unit, whether the category of the first sample is a shared category between the first domain and the second domain; determining, by a second judgment unit, the shared category to which the second sample belongs; and performing joint training on the first judgment unit and the second judgment unit by exchanging information.

[0008] According to yet another aspect of the present disclosure, a computer program capable of implementing the above-mentioned information processing method is also provided.

[0009] In addition, a computer program product in the form of at least a computer-readable storage medium is also provided, on which computer program code for implementing the above-mentioned information processing method is recorded. Specifically, according to another aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer is executed by a computer, the computer program can implement an information processing method for classifying a second sample without a label in a second domain different from the first domain based on a first sample with a label in the first domain, the category of the second sample being a subset or a full set of the category of the first sample, the information processing method comprising: determining by a first judgment unit whether the category of the first sample is a shared category between the first domain and the second domain; determining by a second judgment unit to which the shared category the second sample belongs; and performing joint training on the first judgment unit and the second judgment unit by exchanging information.

[0010] The information processing technology according to the present disclosure proposes an information processing technology based on a new transfer learning model, which can significantly improve the detection accuracy compared with the traditional transfer learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The above and other objects, features and advantages of the present disclosure will be more easily understood with reference to the following description of the embodiments of the present disclosure in conjunction with the accompanying drawings, in which:

[0012] Figure 1 A block diagram illustrating an information processing device according to an embodiment of the present disclosure;

[0013] Figure 2 A schematic diagram illustrating the principle of an information processing device according to an embodiment of the present disclosure;

[0014] Figure 3 A schematic diagram showing a structure of an information processing device according to an embodiment of the present disclosure;

[0015] Figure 4 A schematic diagram illustrating joint training of information processing devices according to an embodiment of the present disclosure;

[0016] Figure 5 A schematic diagram illustrating a training effect of an information processing device according to an embodiment of the present disclosure;

[0017] Figure 6 A flowchart illustrating an information processing method according to an embodiment of the present disclosure; and

[0018] Figure 7 A simplified structural diagram of a general machine for implementing the information processing method according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0019] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to the attached illustrative drawings. When elements of the drawings are indicated by reference numerals, the same elements will be represented by the same reference numerals even though they are shown in different drawings. In addition, in the following description of the present disclosure, detailed descriptions of known functions and configurations incorporated herein will be omitted if they may make the subject matter of the present disclosure unclear.

[0020] The terms used herein are only used for the purpose of describing specific embodiments and are not intended to limit the present disclosure. As used herein, unless the context indicates otherwise, the singular form is intended to also include the plural form. It will also be understood that the terms "comprises," "comprising," and "having" used in the specification are intended to specify the presence of stated features, entities, operations, and / or components, but do not exclude the presence or addition of one or more other features, entities, operations, and / or components.

[0021] Unless otherwise defined, all terms used herein, including technical and scientific terms, have the same meaning as commonly understood by those skilled in the art to which the inventive concepts of the present disclosure belong. 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 art and should not be interpreted in an idealized or overly formal sense unless explicitly defined herein.

[0022] In the following description, many specific details are set forth to provide a comprehensive understanding of the present disclosure. The present disclosure can be implemented without some or all of these specific details. In other examples, in order to avoid obscuring the present disclosure due to unnecessary details, only components closely related to the solutions according to the present disclosure are shown in the accompanying drawings, while other details that are not closely related to the present disclosure are omitted.

[0023] The problem to be solved by the present disclosure is the partial domain adaptation (PDA) problem of transfer learning. In the PDA problem, the source domain (hereinafter also referred to as the first domain) is limited to have labeled samples (hereinafter also referred to as the first samples), and the target domain (hereinafter also referred to as the second domain) has unlabeled samples (hereinafter also referred to as the second samples). In addition, in the PDA problem, the category of the second sample is limited to a subset or a full set of the category of the first sample, that is, the category of the second sample in the target domain is limited to be included in the category of the first sample in the source domain. In addition, in this article, the category of the second sample in the target domain that is included in the category of the first sample in the source domain is called the shared category of the source domain, and the remaining categories within the category of the first sample in the source domain are called the non-shared categories of the source domain. As described above, the number of non-shared categories is a natural number.

[0024] The PDA problem is a problem for classifying unlabeled samples (second samples) of the target domain (second domain) based on labeled samples (first samples) of the source domain (first domain). Specifically, for the PDA problem, assume that the sample set of the source domain is (Sample x i The label, that is, the category is y i ) contains m s samples and n s categories, samples x i The marginal distribution of s , and the target domain sample set Contains m t Sample and n t categories, samples x t The marginal distribution of tIn addition, as mentioned above, n t ≤n s , that is, the categories of the source domain can completely cover the categories of the target domain. In addition, the closure degree of the source domain is defined as ζ = n t / n s The smaller the closure degree ζ, the fewer shared categories there are and the greater the risk of negative transfer.

[0025] In the PDA problem, the main challenge is to align shared categories in the source and target domains. Obviously, the smaller the closure of the source domain, ζ, the greater the probability that samples in the target domain will be matched to non-shared categories, resulting in negative transfer. In the prior art, many methods have been proposed to solve the PDA problem. However, these methods mainly learn domain-invariant features between the source and target domains (e.g., through domain adversarial training) and optimize the classification loss on the source domain to achieve model training in the target domain. However, in the PDA problem, since the source domain contains non-shared categories, the directly extracted domain-invariant features are still inherently misaligned between the two domains (i.e., source and target domains). Therefore, as the closure of the source domain decreases, the domain-invariant features become increasingly inaccurate, resulting in negative transfer. On the other hand, such domain-invariant features may destroy the characteristics unique to the original samples, such as the high similarity between similar samples and the underlying conditional distribution constraints.

[0026] Therefore, one purpose of the information processing technology according to the implementation of the present disclosure is to solve the PDA problem, that is, to train a general model for classifying samples in the target domain without knowing the closure degree of the source domain, so that the general model can have good detection performance within a wider range of closure degrees.

[0027] According to this disclosure, a novel model for solving the PDA problem, called the Coupled Approximation Neural Network (CANN) model, is proposed. Unlike existing methods, the CANN model relaxes the domain-invariant constraint widely used in prior art when solving the PDA problem, instead focusing on mining and leveraging the inherent characteristics of the sample.

[0028] The following combination Figures 1 to 5 The information processing device 100 according to an embodiment of the present disclosure is described in detail.

[0029] Figure 1 A block diagram of the information processing apparatus 100 according to an embodiment of the present disclosure is shown. Figure 2 A schematic diagram illustrating the principle of an information processing device 100 according to an embodiment of the present disclosure. Figure 3A schematic diagram showing the structure of the information processing apparatus 100 according to an embodiment of the present disclosure.

[0030] The information processing device 100 according to an embodiment of the present disclosure is used to classify a second sample without a label in a second domain different from the first domain based on a first sample with a label in the first domain, wherein the category of the second sample is a subset or a full set of the category of the first sample. Figure 1 As shown, the information processing device 100 may include: a first judgment unit 101, configured to judge whether the category of the first sample is a shared category of the first domain and the second domain; and a second judgment unit 102, configured to judge the shared category to which the second sample belongs, wherein the first judgment unit 101 and the second judgment unit 102 perform joint training by exchanging information.

[0031] The information processing device 100 according to an embodiment of the present disclosure may be implemented using the CANN model as described above.

[0032] Specifically, the CANN model according to the present disclosure mines the potential features of samples in both the source domain (i.e., the first domain) and the target domain (i.e., the second domain) based on a self-supervised clustering algorithm. The CANN model according to the present disclosure may include two parallel sub-network models, namely a target domain to source domain approximation neural network (TS ANN) model and a target domain to target domain approximation neural network (TT ANN) model.

[0033] According to an embodiment of the present disclosure, the first judgment unit 101 may be implemented using the TS ANN model described above, and the second judgment unit 102 may be implemented using the TT ANN model described above.

[0034] According to an embodiment of the present disclosure, the first judgment unit 101 and the second judgment unit 102 can be implemented by a convolutional neural network (CNN) model. In addition, according to an embodiment of the present disclosure, the first judgment unit 101 and the second judgment unit 102 can be implemented by a CNN model having the same structure.

[0035] That is, the TS ANN model used to implement the first judgment unit 101 and the TT ANN model used to implement the second judgment unit 102 can both be implemented by CNN models and can have the same model structure.

[0036] In addition, according to an embodiment of the present disclosure, the first sample and the second sample may be images. That is, the sample of the source domain (the first sample) is a labeled image, and the sample of the target domain (the second sample) is an unlabeled image. For example, Figure 3 and 4 As shown in , the source domain can be the domain of cartoon images, and the target domain can be the domain of photo images. Figure 3 and 4 As shown in , for example, an image in the source domain can have four labels, i.e., four categories: bed, chair, schoolbag, TV, etc., while an image in the target domain has only two labels, i.e., two categories: bed and chair.

[0037] As described above, in the PDA problem, the specific category of samples in the target domain, such as photo images, is unknown, but their category is necessarily included in the category of the source domain. The information processing device 100 according to an embodiment of the present disclosure can determine the category of samples in the target domain based on knowledge about samples in the source domain.

[0038] As an example but not a limitation, each of the first judgment unit 101 and the second judgment unit 102 can be implemented using a ResNet-50 model with the same structure that is pre-trained based on the ImageNet image database as a backbone network.

[0039] Since convolutional neural network (CNN) models are well known to those skilled in the art, their details are not described in more detail herein for the sake of brevity.

[0040] Those skilled in the art will recognize that although this document describes exemplary embodiments of the present disclosure using images as samples, the scope of the present disclosure is not limited thereto. Based on the teachings of the present disclosure, those skilled in the art may envision that the information processing apparatus according to the present disclosure may also be applied to other machine learning application fields besides image processing, such as speech analysis, natural language processing, weather forecasting, and the like.

[0041] like Figure 2 As shown in , triangles represent samples in the source domain, while circles represent samples in the target domain. Figure 2 As shown in , the TS ANN model used as the first judgment unit 101 can find source domain neighbor samples corresponding to target domain samples based on feature similarity and implement clustering from the target domain to the source domain using a self-supervised algorithm. Based on the assumption that samples of the same category between two domains are more similar, the TS ANN model can more accurately assess the category differences between the two domains.

[0042] In addition, if Figure 2As shown in , the TT ANN model used as the second judgment unit 102 can find a set of neighboring samples from a target domain sample to other target domain samples based on feature similarity to complete clustering. The TT ANN model can maximize the potential features of the target domain samples while evaluating the reliability (uncertainty) of the target domain samples.

[0043] also, Figure 2 ν and ω shown in will be described in detail below.

[0044] In the information processing apparatus 100 implemented by the CANN model according to the present disclosure, the first judgment unit 101 and the second judgment unit 102 may be implemented as a TS ANN model and a TT ANN model as parallel sub-networks.

[0045] According to an embodiment of the present disclosure, in joint training, the first judgment unit 101 feeds the first information about the shared category (given by Figures 2 to 4 ), and the second judgment unit 102 feeds the second information about the second sample (denoted by Figures 2 to 4 ω is shown in ).

[0046] According to an embodiment of the present disclosure, the first judgment unit 101 and the second judgment unit 102 are not independent of each other, but are jointly trained in a coupled manner. In the forward propagation of the joint training, the TS ANN model used as the first judgment unit 101 can evaluate the category difference between the two domains, while the TT ANN model used as the second judgment unit 102 can evaluate the reliability of the target domain samples. Before the backpropagation of the joint training, the two sub-networks can achieve simultaneous improvement in the performance of the two sub-networks by exchanging information about the category difference between the domains (i.e., the first information ν) and the reliability of the target domain samples (i.e., the second information ω). Receiving information about the category difference between the domains (i.e., the first information ν) from the TS ANN model can help the TT ANN model weaken the impact of negative transfer caused by unshared categories, while receiving information about the reliability of the target domain samples (i.e., the second information ω) from the TT ANN model can help the TS ANN model filter the target domain samples to use more reliable target domain samples to evaluate the category difference between the two domains.

[0047] like Figure 3 As shown, G ts-f and G tt-f G represents the features of samples extracted by the feature extraction layers of the TS ANN model used as the first judgment unit 101 and the TT ANN model used as the second judgment unit 102. ts-y and G tt-yThey respectively represent the classification outputs of the TS ANN model used as the first judgment unit 101 and the TT ANN model used as the second judgment unit 102.

[0048] like Figure 3 As shown, Represents the self-supervised clustering loss of target domain samples to source domain samples, which is used to move target domain samples to corresponding source domain neighbor samples. Represents the self-supervised clustering loss from target domain samples to target domain samples, which is used to move target domain samples to the corresponding target domain neighbor samples. represents the error loss in the TT ANN model used as the second judgment unit 102, which is used to correct the errors of samples classified as non-shared categories in the target domain, thereby eliminating the negative transfer of non-shared categories.

[0049] like Figure 3 As shown, ν may represent the first information about the shared category fed by the TS ANN model serving as the first judgment unit 101 to the TT ANN model serving as the second judgment unit 102. According to an embodiment of the present disclosure, the first information ν may represent the probability that the category of the first sample (e.g., an image in the source domain) is the shared category. Figure 3 As shown in , the probability that the category of the first sample in the source domain is a shared category is shown in the form of a histogram, where the probability that the categories bed and chair are shared categories is the highest.

[0050] According to an embodiment of the present disclosure, can be expressed as the entropy loss combined with ν.

[0051] In addition, if Figure 3 As shown, ω represents the second information about the second sample fed by the TT ANN model used as the second judgment unit 102 to the TS ANN model used as the first judgment unit 101. According to an embodiment of the present disclosure, the second information ω may represent the reliability of the second sample. Figure 3 As shown in , the second samples are shown in order from high reliability to low reliability.

[0052] According to an embodiment of the present disclosure, can be expressed as the motion loss combined with ω.

[0053] According to an embodiment of the present disclosure, the first judgment unit 101 may adjust the second sample in the joint training according to the second information ω received from the second judgment unit 102. In addition, according to an embodiment of the present disclosure, the second judgment unit 102 may adjust the shared category in the joint training according to the first information ν received from the first judgment unit 101.

[0054] Figure 4 A schematic diagram illustrating joint training of the information processing apparatus 100 according to an embodiment of the present disclosure is shown.

[0055] like Figure 4 As shown in FIG, in the joint training, in the TS ANN model used as the first judgment unit 101, F s It can represent the characteristics of the samples extracted from the source domain, and It can represent the characteristics of the sample extracted from the target domain and F s The similarity,

[0056] In addition, for the sample x in the target domain i (1≤i≤m t ), which is different from the sample x in the source domain j (1≤j≤m s ) It can be calculated by the following formula (1).

[0057]

[0058] In formula (1), ||.||2 represents the 2-norm, and T represents the transpose.

[0059] According to the above formula (1), for the sample x in the target domain i , its similarity with samples in the source domain It can be expressed as the following formula (2).

[0060]

[0061] Therefore, the loss function representing the self-supervised clustering loss of target domain samples to source domain samples is It can be calculated by the following formula (3).

[0062]

[0063] In formula (3), D t and D s represent the sample set of the target domain and the sample set of the source domain respectively, and ω i is the second information provided by the TT ANN model used as the second judgment unit 102, which will be described in detail below.

[0064] According to an embodiment of the present disclosure, the TS ANN model used as the first judgment unit 101 can estimate the category difference between the two domains by statistically analyzing the label distribution (i.e., category distribution) of the neighbor samples of the target domain. Specifically, for each category l (1≤l≤n s), which is the weight (probability) ν of the shared category l It can be calculated by the following formula (4).

[0065]

[0066] In formula (4), μ l represents the statistical results of category l in the source domain, For the sample x in the target domain i , as the accumulation weight μ l It can be updated by the following formula (5).

[0067]

[0068] In formula (5),

[0069] According to an embodiment of the present disclosure, after traversing all target domain samples each time, the first information ν will be reset.

[0070] In addition, if Figure 4 As shown in FIG, in the joint training, in the TT ANN model used as the second judgment unit 102, F t It can represent the characteristics of the sample extracted from the target domain, and It can represent the characteristics of the sample extracted from the target domain and F t Here, because when calculating the neighbor samples of the target domain, it is necessary to consider excluding the samples of the target domain itself, so

[0071] In addition, for the sample x in the target domain i (1≤i≤m t ), which is different from other samples x in the target domain j (1≤j≤m t , j≠i) It can be calculated by the following formula (6).

[0072]

[0073] Therefore, for a sample x in the target domain i , its similarity with other samples in the target domain It can be expressed as the following formula (7).

[0074]

[0075] Therefore, the loss function representing the self-supervised clustering loss of target domain samples to target domain samples is It can be calculated by the following formula (8).

[0076]

[0077] In the estimated target domain sample x i When the reliability of the target domain is calculated, it can be noted that the samples of the target domain that are more difficult to classify or have lower reliability are more likely to be isolated in the feature space. Therefore, according to the embodiment of the present disclosure, the TT ANN model used as the second judgment unit 102 can be used for the samples x of the target domain. i Calculate the Kullback-Leibler divergence (ie, relative entropy) of its similarity matrix relative to the uniform distribution to estimate the sample x of the target domain i The reliability of this can be determined by the second information ω i It can be expressed as follows:

[0078]

[0079] In equation (9), θ can be taken as a minimum value to avoid log0.

[0080] Thus, the error loss in the TT ANN model used as the second judgment unit 102 is It can be calculated by the following formula (10).

[0081]

[0082] In formula (10), where y i,j represents a sample x of the target domain output by the TT ANN model used as the second judgment unit 102 i The probability that the label (category) is j.

[0083] According to the embodiments of the present disclosure, considering the constraints of hardware computing power, the training of the CANN model of the information processing device 100 (including the TS ANN model used as the first judgment unit 101 to the TT ANN model used as the second judgment unit 102) can be carried out in batches. Figure 4 As shown in , it is assumed that the number of samples of the target domain in each batch is m. For each batch, the above formulas (1) to (10) are also applicable, and only m needs to be replaced by t Just replace it with m.

[0084] Using the loss functions given by equations (3), (8) and (10) above and The TS ANN model and the TT ANN model included in the CANN model may be jointly trained to obtain the final information processing device 100 .

[0085] like Figure 4 As shown in , in joint training, the TS ANN model used as the first judgment unit 101 feeds first information ν about the shared category to the TT ANN model used as the second judgment unit 102, and the TT ANN model used as the second judgment unit 102 feeds second information ω about the samples of the target domain to the TS ANN model used as the first judgment unit 101.

[0086] As mentioned above, the higher the reliability of the sample in the target domain, the easier it is to be correctly matched to the shared category. Conversely, the lower the reliability of the sample in the target domain, the more difficult it is to be matched to the shared category, that is, the greater the uncertainty. According to the embodiment of the present disclosure, the second information ω can reflect the reliability (uncertainty) of the sample in the target domain. In this regard, Figure 4 As shown in , during joint training, the TS ANN model serving as the first judgment unit 101 can utilize the second information ω to amplify the reliability gap between highly reliable samples and less reliable samples, thereby controlling the concentration of the sample distribution. In other words, the second information ω makes the distribution of reliable target domain samples more concentrated (sharp) while making the distribution of unreliable target domain samples more dispersed (slow), thereby obtaining a more deterministic result.

[0087] In addition, in order to reduce the negative transfer effect of non-shared categories in the target domain, in the joint training, according to an embodiment of the present disclosure, the TT ANN model used as the second judgment unit 102 uses the first information ν from the TS ANN model used as the first judgment unit 101. According to an embodiment of the present disclosure, the first information ν can reflect the probability (weight) that a certain category in the source domain is a shared category. In this regard, Figure 4 As shown in , in joint training, the TT ANN model used as the second judgment unit 102 can use the first information ν to increase the probability (entropy loss) of the shared category and reduce the probability (entropy loss) of the non-shared category, so as to weaken the negative migration effect of the non-shared category in the target domain.

[0088] It should be understood that the first information ν and the second information ω exchanged with each other during the joint training process of the TS ANN model used as the first judgment unit 101 and the T-TANN model used as the second judgment unit 102 can achieve simultaneous improvement in the performance of the TS ANN model and the TT ANN model.

[0089] The information processing device 100 according to the embodiment of the present disclosure is implemented by the CANN model, which achieves a significant improvement in accuracy when solving the PDA problem compared to the prior art methods.

[0090] For example, the experiment was based on the Office-Home image database commonly used in the prior art. It includes four domains: Art (Ar), Clipart (Cl), Product (Pr), and Real World (Rw). Each domain contains 65 categories. For each domain, the first 25 categories in alphabetical order are used as shared categories, and the remaining categories are used as non-shared categories. Information processing device 100 according to an embodiment of the present disclosure is trained based on this database.

[0091] Experimental results show that the information processing device 100 according to the embodiment of the present disclosure can achieve an average detection accuracy of 75.5% for samples in the target domain, which is at least 4% higher than the existing method.

[0092] Figure 5 A schematic diagram illustrating a training effect of an information processing device according to an embodiment of the present disclosure.

[0093] like Figure 5 As shown in (a), the three curves from top to bottom show and As the joint training iterations proceed, the loss tends to be stable. Figure 5 As shown in (b), the weights of the 65 categories in the domain are shared categories. Through joint training, the shared categories (such as the top 25 categories) and non-shared categories in the source domain can be determined more accurately.

[0094] Accordingly, the present disclosure also proposes an information processing method for transfer learning between different domains.

[0095] Figure 6 is a flowchart illustrating an information processing method 600 according to an embodiment of the present disclosure.

[0096] The information processing method 600 starts at step S601. Then, in step S602, the first judgment unit judges whether the category of the first sample is a shared category between the first domain and the second domain. According to the embodiment of the present disclosure, the processing in step S602 can be performed, for example, according to the above reference. Figures 1 to 5 The first judgment unit 101 described above is implemented, so its details are not repeated here.

[0097] Then, in step S603, the second judgment unit determines the shared category to which the second sample belongs. According to the embodiment of the present disclosure, the processing in step S603 can be performed, for example, according to the above reference. Figures 1 to 5 The second judgment unit 102 described above is implemented, so its details are not repeated here.

[0098] Subsequently, in step S604, joint training is performed on the first judgment unit and the second judgment unit by exchanging information. According to an embodiment of the present disclosure, the processing in step S604 can be performed, for example, by referring to the above reference. Figures 1 to 5 The interaction between the first judgment unit 101 and the second judgment unit 102 is described, and thus its details are not repeated here.

[0099] Finally, the information processing method 600 ends at step S605.

[0100] According to the information processing technology of the present disclosure, it is possible to achieve a significant improvement in the detection accuracy of transfer learning without destroying the unique features of the original samples.

[0101] Figure 7 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. General-purpose machine 700 can be, for example, a computer system. It should be noted that general-purpose machine 700 is merely an example and does not imply any limitation on the scope of use or functionality of the information processing method and information processing apparatus of the present disclosure. Furthermore, general-purpose machine 700 should not be interpreted as having any dependency or requirement on any of the components or combinations thereof shown in the aforementioned information processing method or information processing apparatus.

[0102] exist Figure 7 In the embodiment of the present invention, a central processing unit (CPU) 701 executes various processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage section 708 to a random access memory (RAM) 703. In the RAM 703, images required when the CPU 701 executes various processes, etc., are also stored as needed. The CPU 701, the ROM 702, and the RAM 703 are connected to each other via a bus 704. An input / output interface 705 is also connected to the bus 704.

[0103] The following components are also connected to the input / output interface 705: an input section 706 (including a keyboard, a mouse, etc.), an output section 707 (including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.), a storage section 708 (including a hard disk, etc.), and a communication section 709 (including a network interface card such as a LAN card, a modem, etc.). The communication section 709 performs communication processing via a network such as the Internet. A drive 710 may also be connected to the input / output interface 705 as needed. A removable medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. can be installed in the drive 710 as needed, so that a computer program read therefrom can be installed in the storage section 708 as needed.

[0104] In the case of realizing the above-described series of processing by software, a program constituting the software can be installed from a network such as the Internet or from a storage medium such as the removable medium 711 .

[0105] It should be understood by those skilled in the art that such storage media is not limited to Figure 7 The removable medium 711 shown has a program stored therein and is distributed separately from the device to provide the program to the user. Examples of the removable medium 711 include magnetic disks (including floppy disks), optical disks (including compact disk read-only memories (CD-ROMs) and digital versatile disks (DVDs)), magneto-optical disks (including minidiscs (MD) (registered trademark)), and semiconductor memories. Alternatively, the storage medium may be ROM 702, a hard disk included in storage section 708, or the like, in which the program is stored and distributed to the user together with the device containing it.

[0106] In addition, the present disclosure also provides a program product storing machine-readable instruction code. When the instruction code is read and executed by a machine, the information processing method according to the present disclosure is executed. Accordingly, the various storage media listed above for carrying such a program product are also included within the scope of the present disclosure.

[0107] The above has been described in detail through block diagrams, flow charts and / or embodiments to illustrate specific embodiments of the apparatus and / or method according to the embodiments of the present disclosure. When these block diagrams, flow charts and / or embodiments include one or more functions and / or operations, those skilled in the art will understand that each function and / or operation in these block diagrams, flow charts and / or embodiments can be implemented individually and / or collectively through various hardware, software, firmware or any combination thereof. In one embodiment, several portions of the subject matter described in this specification can be implemented through 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 equivalently implemented in whole or in part in the form of one or more computer programs running on one or more computers (for example, in the form of one or more computer programs running on one or more computer systems), in the form of one or more programs running on one or more processors (for example, in the form of one or more programs running on one or more microprocessors), in the form of firmware, or in substantially any combination thereof, and that, based on the contents disclosed in this specification, it is well within the capabilities of those skilled in the art to design circuits for the present disclosure and / or write code for software and / or firmware for the present disclosure.

[0108] In summary, the following technical solutions are provided in the embodiments according to the present disclosure, but are not limited to these technical solutions:

[0109] Solution 1. An information processing device for classifying, based on first labeled first samples from a first domain, second unlabeled samples from a second domain different from the first domain, wherein the category of the second sample is a subset or the full set of the category of the first sample, the information processing device comprising:

[0110] a first determining unit configured to determine whether the category of the first sample is a shared category between the first domain and the second domain; and

[0111] A second determining unit is configured to determine the sharing category to which the second sample belongs,

[0112] The first judgment unit and the second judgment unit perform joint training by exchanging information.

[0113] Solution 2. An information processing device according to Solution 1, wherein the first judgment unit and the second judgment unit are implemented by a convolutional neural network model.

[0114] Solution 3. An information processing device according to Solution 2, wherein the first judgment unit and the second judgment unit are implemented by a convolutional neural network model having the same structure.

[0115] Solution 4. The information processing device according to solution 1, wherein the first sample and the second sample are images.

[0116] Option 5. An information processing device according to any one of Options 1 to 4, wherein, in the joint training, the first judgment unit feeds first information about the shared category to the second judgment unit, and the second judgment unit feeds second information about the second sample to the first judgment unit.

[0117] Solution 6. The information processing device according to Solution 5, wherein the first information indicates a probability that the category of the first sample is the shared category.

[0118] Solution 7. The information processing device according to Solution 5, wherein the second information indicates the reliability of the second sample.

[0119] Solution 8. The information processing device according to Solution 5, wherein the first judgment unit adjusts the second sample in the joint training according to the second information received from the second judgment unit.

[0120] Solution 9. The information processing device according to Solution 5, wherein the second judgment unit adjusts the shared category in the joint training according to the first information received from the first judgment unit.

[0121] Solution 10. An information processing method for classifying, based on first labeled first samples in a first domain, second unlabeled samples in a second domain different from the first domain, wherein the category of the second sample is a subset or the full set of the category of the first sample, the information processing method comprising:

[0122] determining, by a first judging unit, whether the category of the first sample is a shared category between the first domain and the second domain;

[0123] Determining, by a second determining unit, the sharing category to which the second sample belongs; and

[0124] Joint training is performed on the first judgment unit and the second judgment unit by exchanging information.

[0125] Solution 11. The information processing method according to Solution 10, wherein the first judgment unit and the second judgment unit are implemented by a convolutional neural network model.

[0126] Solution 12. The information processing method according to Solution 11, wherein the first judgment unit and the second judgment unit are implemented by a convolutional neural network model having the same structure.

[0127] Solution 13. The information processing method according to Solution 10, wherein the first sample and the second sample are images.

[0128] Scheme 14. An information processing method according to any one of Schemes 10 to 13, wherein, in the joint training, the first judgment unit feeds first information about the shared category to the second judgment unit, and the second judgment unit feeds second information about the second sample to the first judgment unit.

[0129] Solution 15. The information processing method according to Solution 14, wherein the first information represents a probability that the category of the first sample is the shared category.

[0130] Solution 16. The information processing method according to Solution 14, wherein the second information represents the reliability of the second sample.

[0131] Solution 17. The information processing method according to Solution 14, wherein, in the joint training, the first judgment unit adjusts the second sample according to the second information received from the second judgment unit.

[0132] Solution 18. The information processing method according to Solution 14, wherein, in the joint training, the second judgment unit adjusts the shared category according to the first information received from the first judgment unit.

[0133] Solution 19. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a computer, can implement the information processing method according to any one of Solutions 10 to 18.

[0134] Although the present disclosure has been disclosed above through the description of specific embodiments of the present disclosure, it should be understood that those skilled in the art may design various modifications, improvements, or equivalents to the embodiments of the present disclosure within the spirit and scope of the appended claims. Such modifications, improvements, or equivalents should also be considered to be included within the scope of protection of the present disclosure.

Claims

1. An information processing apparatus for classifying, based on first labeled first samples from a first domain, second unlabeled samples from a second domain different from the first domain, wherein the categories of the second samples are a subset or a full set of the categories of the first samples, the information processing apparatus comprising: a first determining unit configured to determine whether the category of the first sample is a shared category between the first domain and the second domain; as well as A second determining unit is configured to determine the sharing category to which the second sample belongs, The first judgment unit and the second judgment unit perform joint training by exchanging information. wherein the first sample and the second sample are images, and In the joint training, the first judgment unit feeds first information about the shared category to the second judgment unit, and the second judgment unit feeds second information about the second sample to the first judgment unit.

2. The information processing device according to claim 1, wherein The first judgment unit and the second judgment unit are implemented by a convolutional neural network model.

3. The information processing device according to claim 2, wherein: The first judgment unit and the second judgment unit are implemented by a convolutional neural network model with the same structure. The information processing device according to claim 1 , wherein: The first information indicates a probability that the category of the first sample is the shared category. The information processing device according to claim 1 , wherein: The second information indicates the reliability of the second sample. The information processing apparatus according to claim 1 , wherein: The first judgment unit adjusts the second sample in the joint training according to the second information received from the second judgment unit.

7. The information processing apparatus according to claim 1, wherein: The second judgment unit adjusts the shared category in the joint training according to the first information received from the first judgment unit.

8. An information processing method for classifying, based on first labeled first samples from a first domain, second unlabeled samples from a second domain different from the first domain, wherein the category of the second sample is a subset or the full set of the category of the first sample, the information processing method comprising: determining, by a first judging unit, whether the category of the first sample is a shared category between the first domain and the second domain; Determining, by a second determining unit, the sharing category to which the second sample belongs; as well as Performing joint training on the first judgment unit and the second judgment unit by exchanging information, wherein the first sample and the second sample are images, and In the joint training, the first judgment unit feeds first information about the shared category to the second judgment unit, and the second judgment unit feeds second information about the second sample to the first judgment unit.

Images