Information processing apparatus, information processing method, and machine-readable storage medium

Abstract

The present disclosure relates to an information processing apparatus, an information processing method, and a machine-readable storage medium. The information processing apparatus includes a feature generation unit configured to generate features for input distributions, a difference generation unit configured to generate output feature differences of the feature generation unit trained well at a predetermined stage and the feature generation unit trained at a current stage for the same specific input distributions, and an update unit configured to train the feature generation unit of the current stage using output features of the input distributions online updated at the current stage by the feature generation unit trained at the current stage based on the output feature differences. The information processing apparatus can train the feature generation unit using the output feature differences of the same input distributions by the feature generation unit, thereby effectively solving the forgetting problem in online unsupervised domain adaptation.

Description

Technical Field

[0001] This disclosure relates to the field of information processing, and more particularly to an information processing apparatus, information processing method, and machine-readable storage medium for unsupervised domain adaptation. Background Technology

[0002] This section provides background information relating to this disclosure, which is not necessarily prior art.

[0003] Figure 1 This diagram compares several machine learning methods, where the horizontal axis represents labeling cost and the vertical axis represents model performance. For example... Figure 1 As shown, while traditional supervised learning offers high performance, it requires extensive manual annotation of the target domain dataset, resulting in high costs. To reduce annotation costs, unsupervised domain adaptation (UDA) learning has become a popular research direction. Unsupervised domain adaptation involves two subsets of datasets with the same task: a labeled source domain dataset and an unlabeled target domain dataset. The goal of unsupervised domain adaptation is to leverage the labeled source domain dataset to improve the model's performance on the target domain dataset, thereby reducing the annotation costs on the target domain dataset.

[0004] Unsupervised Domain Adaptation (UDA) has been widely applied in many fields, such as handwriting recognition, image classification, image segmentation, and object detection. However, current unsupervised domain adaptation techniques are based on offline training, but offline UDA has the following limitations: it requires the pre-collection of a large amount of target domain dataset, resulting in weak cold-start capability; it requires a long time to complete model training; and it is not sensitive to data changes, requiring retraining the model when the data changes, leading to significant overhead.

[0005] Therefore, offline training methods cannot meet the needs of most practical applications, especially in real-time systems. In real-world applications, users typically collect data in the form of real-time data streams and expect to update the model online in real time, i.e., updating the model in real time based on mini-batch target domain data. Therefore, this application will discuss Online Unsupervised Domain Adaptation (OUDA), which is an unsupervised online real-time update of the model based on real-time data streams. Summary of the Invention

[0006] This section provides a general overview of this disclosure, rather than a full disclosure of its entire scope or all its features.

[0007] The purpose of this disclosure is to provide an adaptive information processing apparatus, information processing method, and machine-readable storage medium for online unsupervised domains.

[0008] According to one aspect of this disclosure, an information processing apparatus for online unsupervised domain adaptation is provided, comprising: a feature generation unit configured to generate features for an input distribution; a difference generation unit configured to generate output feature differences between a feature generation unit trained in a predetermined stage and a feature generation unit trained in the current stage for the same specific input distribution; and an update unit configured to train a feature generation unit in the current stage based on the output feature differences using the output features of the feature generation unit trained in the current stage for an input distribution updated online in the current stage.

[0009] According to another aspect of this disclosure, a method for adaptive information processing in an online unsupervised domain is provided, comprising: generating output feature differences between a feature generation unit trained in a predetermined stage and a feature generation unit trained in the current stage for the same specific input distribution; and training a feature generation unit in the current stage based on the output feature differences using the output features of the feature generation unit trained in the current stage for an input distribution updated online in the current stage.

[0010] According to another aspect of this disclosure, a machine-readable storage medium is provided that carries a program product including machine-readable instruction code stored thereon, wherein the instruction code, when read and executed by a computer, enables the computer to perform an information processing method according to this disclosure.

[0011] Using the information processing apparatus, information processing method, and machine-readable storage medium according to the present disclosure, a feature generation unit can be trained by utilizing the differences in output features of the same input distribution, thereby effectively solving the forgetting problem in online unsupervised domain adaptation.

[0012] Further applicability will become apparent from the description provided herein. The descriptions and specific examples in this summary are for illustrative purposes only and are not intended to limit the scope of this disclosure. Attached Figure Description

[0013] The accompanying drawings described herein are for illustrative purposes only and not for all possible implementations, and are not intended to limit the scope of this disclosure. In the drawings:

[0014] Figure 1 The diagram illustrates a comparison of several machine learning methods;

[0015] Figure 2 This is a schematic diagram illustrating the principle of OUDA according to this disclosure;

[0016] Figure 3 This is a block diagram illustrating the structure of an information processing apparatus according to an embodiment of the present disclosure;

[0017] Figure 4 This is a block diagram illustrating the structure of an information processing apparatus according to another embodiment of the present disclosure;

[0018] Figure 5 This is a schematic diagram illustrating the principle of an information processing apparatus according to another embodiment of the present disclosure;

[0019] Figure 6 This is a rendering of an information processing apparatus according to an embodiment of the present disclosure;

[0020] Figure 7 This is a diagram illustrating the effect of an information processing apparatus according to an embodiment of the present disclosure in resisting forgetting.

[0021] Figure 8 This is a flowchart illustrating an information processing method according to an embodiment of the present disclosure;

[0022] Figure 9 A flowchart illustrating an information processing method according to another embodiment of the present disclosure; and

[0023] Figure 10 This is a block diagram illustrating an exemplary structure of a general-purpose personal computer in which information processing apparatus and methods according to embodiments of the present disclosure can be implemented.

[0024] While this disclosure is readily subject to various modifications and substitutions, specific embodiments thereof have been shown by way of example in the accompanying drawings and are described in detail herein. However, it should be understood that the description of specific embodiments herein is not intended to limit this disclosure to the specific forms disclosed, but rather, this disclosure is intended to cover all modifications, equivalents, and substitutions falling within the spirit and scope of this disclosure. It should be noted that throughout the drawings, corresponding reference numerals indicate corresponding parts. Detailed Implementation

[0025] Examples of this disclosure will now be described more fully with reference to the accompanying drawings. The following description is merely exemplary and is not intended to limit the disclosure, its application, or its uses.

[0026] Example embodiments are provided so that this disclosure will be thorough, and will fully convey the scope thereof to those skilled in the art. Numerous specific details are set forth such as examples of specific components, devices, and methods, to provide a thorough understanding of embodiments of the present disclosure. It will be apparent to those skilled in the art that specific details need not be employed, that the example embodiments may be embodied in many different forms and that neither should be construed to limit the scope of the present disclosure. In some example embodiments, well-known processes, well-known structures, and well-known technologies have not been described in detail.

[0027] Similar to UDA, OUDA contains two sub-datasets with the same task: a labeled source domain dataset and an unlabeled target domain dataset. The goal of OUDA is to use the labeled source domain dataset to improve the performance of the model on the online-updated target domain dataset, thereby reducing the annotation cost on the target domain dataset and achieving real-time updates of the model. OUDA consists of consecutive learning phases. For each phase, the algorithm inherits the model generated in the previous phase and updates the model using the currently obtained unlabeled target domain dataset. A key challenge of OUDA is to relate new knowledge to old knowledge, i.e., to learn while resisting forgetting.

[0028] For ease of understanding, as an example, the task of OUDA can be to distinguish four types of objects: beds, chairs, bicycles, and forks. The source domain dataset can be images of real-life scenes of beds, chairs, bicycles, and forks, while the target domain dataset can be images of a cartoon style of beds, chairs, bicycles, and forks. Since the source domain dataset has annotations, a source model with good performance can be obtained for the source domain dataset. However, the target domain dataset is unannotated, and since the dataset style of the target domain dataset is different from that of the source domain dataset, the source model is not effectively available for the target domain dataset either. OUDA expects to improve the performance of the target domain dataset by transferring knowledge from the labeled source domain dataset to the unlabeled target domain dataset.

[0029] In OUDA, it is assumed that the target domain dataset is obtained in small batches in sequence, and the model is transferred only for the small batch of the target domain dataset each time. In this way, the training time for each time is relatively short, so that real-time updates of the model can be achieved in an online form. Here, a small batch can represent a dataset including several to dozens of data samples. And, the model in this article refers to the model of the feature generator (hereinafter also referred to as the feature generation unit). In addition, OUDA in this article is implemented using a neural network.

[0030] <The forgetting problem in OUDA>

[0031] Unlike offline learning, online model updates can lead to a severe forgetting problem. That is, updating the model based on newly collected target domain data can degrade its performance on previously collected target domain data, resulting in poor generalization performance and model training failure to converge. First, refer to... Figure 6 Let me briefly describe the problem of forgetting.

[0032] exist Figure 6 In the diagram, the horizontal axis represents the number of stages in online training (i.e., the number of batches of the target domain dataset added in batches), and the vertical axis represents the classification accuracy of the entire target domain in the current stage. Figure 6 In the example, 36 target samples are obtained at each stage. Furthermore, the dots represent the performance of the classic unsupervised domain adaptation algorithms DANN and CDAN during online training; upward-pointing triangles represent the performance after incorporating the loss function according to this application; and downward-pointing triangles represent the performance when directly applying the source-only model to the target domain. This experiment is based on the Office-31 dataset. Office-31 is a widely used public dataset for evaluating transfer learning algorithms. Figure 6 As shown by the dotted lines, DANN and CDAN exhibit significant fluctuations in performance in the target domain during this task, resulting in catastrophic forgetting, i.e., poor generalization and convergence. Therefore, traditional offline training methods are not applicable to online learning tasks.

[0033] <Measures for measuring forgetting>

[0034] This application proposes a modeling approach for the forgetting problem in online unsupervised domain adaptation and introduces a metric for measuring forgetting: output feature difference. This metric primarily evaluates the difference in output distribution across different stages of the model for the same input distribution during continuous online learning. A small difference indicates good resistance to forgetting across different stages, while a large difference indicates strong resistance to forgetting. To resist forgetting, more constraints are needed on model updates.

[0035] The feature generator g can be applied to the same input distribution. The output feature differences are represented as follows:

[0036]

[0037] In equation (1), g 1 and g 2 These represent the feature generator models trained at different stages, and x represents the input distribution. The data in g 1 (x) and g 2 (x) represent g respectively 1 and g2 Regarding the output features of data x, and They represent g respectively 1 (x) and g 2 The distribution of (x), This represents finding the supremum of the difference in output distributions between two feature generators on any common input distribution.

[0038] To minimize the difference in output features between the current model and previous models, feature shift between the two models can be minimized by using the widest possible common input distribution. For example, since the target domain dataset is used to acquire new knowledge, constraints can be imposed on stable source domain features to reduce output feature differences. In this case, the constraints do not weaken the model's learning ability in the target domain. Figure 2 As shown, with Figure 1 Compared to the UDA framework in the previous application, the OUDA according to this application can utilize the source model to impose constraints on the current model at the current stage to resist forgetting.

[0039] <Principles and Structure of Information Processing Devices>

[0040] The following is for reference. Figure 3 To illustrate the structure of the information processing apparatus according to this application. For example... Figure 3 As shown, the information processing apparatus 100 according to this application may include a feature generation unit 110, a difference generation unit 120, and an update unit 130.

[0041] The feature generation unit 110 can generate corresponding features for the input distribution.

[0042] Furthermore, the difference generation unit 120 can generate the output feature difference between the feature generation unit trained in the predetermined stage and the feature generation unit trained in the current stage for the same specific input distribution, and provide the obtained output feature difference to the update unit 130.

[0043] Specifically, the same specific input distribution is the same labeled source domain dataset. The feature generation unit trained in the predetermined stage is the feature generation unit trained with only the source domain dataset as the input distribution, and the feature generation unit trained in the current stage is the feature generation unit trained with the input distribution based on the source domain dataset, and the target domain dataset is added in stages, where the current stage is any one of multiple stages of adding the target domain dataset.

[0044] Furthermore, the update unit 130 can train the feature generation unit of the current stage based on the output feature differences generated by the difference generation unit 120, using the output features of the feature generation unit trained in the current stage for the input distribution updated online in the current stage.

[0045] Therefore, the information processing apparatus 100 according to this application can use the difference in output features of the feature generation unit on the same input distribution to train the feature generation unit, thereby effectively solving the forgetting problem in OUDA.

[0046] The following is combined with Figure 4 and Figure 5 The structure and principle of an information processing apparatus 200 according to another embodiment of this application will be described.

[0047] As mentioned above, to address the forgetting problem in online learning, it is necessary to minimize the differences in output features while preserving the model's learning ability on the target domain dataset. Therefore, this application proposes Source Feature Distillation (SFD). SFD uses an additional source model (a convergent model trained on source domain data) to constrain the target model, thereby learning source domain features that will not be forgotten. SFD does not constrain the memorability of target domain data features, thus maximizing the preservation of the model's learning ability on the target domain data.

[0048] like Figure 4 The information processing device 200 may include a feature generation unit 210, a difference generation unit 220, an update unit 230, and a loss function generation unit 240, wherein the update unit 230 may include a training unit 2301, a classification unit 2302, and a domain discrimination unit 2303. The feature generation unit 210, the difference generation unit 220, and the update unit 230 correspond to... Figure 3 The feature generation unit 110, difference generation unit 120, and update unit 130 are included.

[0049] The feature generation unit 210 can generate corresponding features for the input distribution.

[0050] Furthermore, the difference generation unit 220 can generate the output feature difference between the feature generation unit trained in a predetermined stage and the feature generation unit trained in the current stage for the same specific input distribution.

[0051] Specifically, the same specific input distribution is the same labeled source domain dataset. The feature generation unit trained in the predetermined stage is the feature generation unit trained with only the source domain dataset as the input distribution, and the feature generation unit trained in the current stage is the feature generation unit trained with the input distribution based on the source domain dataset, and the target domain dataset is added in stages, where the current stage is any one of multiple stages of adding the target domain dataset.

[0052] Furthermore, the loss function generation unit 240 can generate a loss function to represent the differences in output features and provide it to the update unit 230. It should be noted that the difference generation unit 220 and the loss function generation unit 240 can also be merged, meaning that a loss function representing the differences in output features can be generated directly.

[0053] As an example, the loss function is the loss function of the output features of the feature generation unit in the current stage for the source domain dataset and the output features of the feature generation unit for the source domain dataset when the input distribution only contains the source domain dataset.

[0054] As another example, the loss function is the sum of the loss functions of the output features of the feature generation unit in the current stage for the source domain dataset and the output features of the feature generation units in all previous stages for the source domain dataset.

[0055] Furthermore, the loss function mentioned above can be any distance loss function, such as L2 distance loss function, L1 distance loss function, cosine distance loss function, etc.

[0056] Furthermore, the update unit 230 can train the feature generation unit of the current stage based on the loss function using the output features of the feature generation unit trained in the current stage for the input distribution updated online in the current stage.

[0057] Specifically, the training unit 2301 in the update unit 230 can train the feature generation unit of the current stage based on the loss function using the analysis results of the output features of the feature generation unit of the current stage for the input distribution of the current stage, so as to update the feature generation unit in real time.

[0058] The classification unit 2302 in the update unit 230 can classify the objects in the input distribution of the current stage based on the output features of the feature generation unit of the current stage for the input distribution of the current stage, and provide the classification results to the training unit 2301 for training the feature generation unit.

[0059] The domain discrimination unit 2303 in the update unit 230 can classify whether the input distribution of the current stage belongs to the source domain or the target domain based on the output features of the current stage feature generation unit for the current stage input distribution, and provide the classification results to the training unit 2301 for training the feature generation unit.

[0060] In addition, such as Figure 4 As shown, Figure 4 The domain discrimination unit 2303 is indicated by a dashed box, which means that the domain discrimination unit 2303 is not mandatory. Instead of the domain discrimination unit 2303, or in addition to the domain discrimination unit 2303, the update unit 230 may also have additional units for analyzing the mixing of the source and target domains.

[0061] The following is combined with Figure 5 To describe in more detail Figure 4 The information processing device 200 in the middle processes the information. Figure 5 This is a schematic diagram illustrating the principle of an information processing apparatus according to another embodiment of the present disclosure. For ease of description, in Figure 5 Some units of the information processing device 200 are omitted. Figure 5 In the diagram, x represents the data in the input distribution (where x... s Represents the data in the source domain, x t This represents the data in the target domain at the current stage t (where t is a natural number greater than or equal to 1), g t g represents the feature generation unit (model) to be learned in the current stage t. 0 It is the feature generation unit (model) when only the source domain dataset is available, and z represents the output feature space generated by the feature generation unit.

[0062] Since only the source domain dataset has labels, classification unit 2302 classifies objects in the source domain dataset (e.g., images) based on the output features of the source domain, and then provides the classification results to training unit 2301. Training unit 2301 can train the feature generation unit based on the true labels of the source domain dataset and the classification results of classification unit 2302.

[0063] Furthermore, the training unit 2301 can also train the feature generation unit based on the discrimination results of the domain discrimination unit 2303 on the source and target domains, so that the source and target domains can be mixed together and are not easily distinguished. In other words, the training unit 2301 aims to obtain domain-invariant features, that is, features that are effective for both the source and target domains.

[0064] As an example, suppose the task is to distinguish between cats and dogs, and the images in the source domain are images of cats and dogs in real-world scenes, while the images in the target domain are images of cats and dogs in cartoon style. For instance, with only the source domain dataset, the feature generation unit might learn to distinguish between cats and dogs using color (for simplicity, as an example). However, with the addition of the target domain dataset, color might become less effective, and the feature generation unit might learn to better distinguish between cats and dogs in the source and target domains using the leg length of cats and dogs (for simplicity, as an example). Therefore, in this example, the domain-invariant feature that training unit 2301 might learn is the leg length of cats and dogs.

[0065] The following will describe Figure 5 To address the forgetting problem in OUDA, a constraint, i.e., a loss function, is applied to source domain features to ensure stable performance. loss function It can be done Figure 4 The loss function is generated by the loss function generation unit 240.

[0066] The total loss of OUDA in the current stage t can be expressed as:

[0067]

[0068] In equation (2), Let represent the loss function of UDA, and This represents the data distillation loss function of the source domain, where α represents the loss function used for balancing. and The coefficient.

[0069] in, It can be represented as:

[0070]

[0071] In equation (3), g t It is the feature generation unit (model) that t needs to learn at the current stage, g i It is the feature generation unit (model) in the previous stage i before the current stage, i∈t-1, symbol To express summation, the symbol is... Let D represent the loss functions of A and B, while D represents the loss functions of B and C. s This represents the dataset (data distribution) of the source domain.

[0072] To simplify The computation can use a model with only the source domain as the initialization model and... Represented as:

[0073]

[0074] Here, it is important to note the loss function. It can be any distance loss function, such as L2 distance loss function, L1 distance loss function, cosine distance loss function, etc.

[0075] As an example, in the loss function When using the L2 distance loss function, we can obtain:

[0076]

[0077] From the above As can be seen from the formula, the loss function It represents the difference in output features between the current model and the source model of the feature generation unit at the current stage t.

[0078] Refer again Figure 6 ,from Figure 6 As can be seen from the lines connecting the upward triangles, the forgetting problem of OUDA is greatly improved compared to the case where SFD is not applied, represented by the lines connecting the dots.

[0079] In addition, such as Figure 5 As shown, the SFD proposed in this application can be viewed as a plug-in method. It is worth noting that the proposed SFD is not tightly integrated with domain adversarial training methods. Therefore, SFD can be applied to any OUDA method for learning domain-invariant features, i.e., introducing a single feature generator with similar distributions in two domains. This can also be seen from... Figure 6 It can be seen that SFD can solve the forgetting problem of OUDA, whether for CDAN or DANN.

[0080] Figure 7 This is a diagram illustrating the effect of an information processing apparatus according to an embodiment of the present disclosure in resisting forgetting. Figure 7 The horizontal axis represents the number of stages, and the vertical axis represents the average forgetting rate in the target domain. The higher the value of the vertical axis, the higher the forgetting rate. Figure 7 The experiments on resistance to forgetting used three benchmark algorithms: DANN, CDAN, and MCC. Light gray represents the experimental results without SFD, while dark gray represents the results with SFD. The experimental results show that in the target domain, using SFD can reduce forgetting by approximately 70%.

[0081] Therefore, the information processing apparatus 200 according to another embodiment of the present disclosure can use an SFD-based loss function to minimize the differences in output features while preserving the model's learning ability on the target domain dataset, thereby solving the forgetting problem of OUDA and effectively improving the generalization and convergence of the model in online learning.

[0082] Information Processing Methods

[0083] The following is combined with Figure 8 This describes an information processing method according to embodiments of the present disclosure.

[0084] like Figure 8 As shown, the information processing method according to an embodiment of this disclosure begins at step S110. In step S110, the output feature difference between a feature generation unit trained in a predetermined stage and a feature generation unit trained in the current stage for the same specific input distribution is generated.

[0085] Specifically, the same specific input distribution belongs to the same labeled source domain dataset. The feature generation unit trained in the predetermined stage is the feature generation unit trained with only the source domain dataset as the input distribution, and the feature generation unit trained in the current stage is the feature generation unit trained with the input distribution based on the source domain dataset, and the target domain dataset is added in stages, where the current stage is any one of multiple stages of adding the target domain dataset.

[0086] Next, in step S120, the feature generation unit for the current stage is trained using the output features of the feature generation unit trained in the current stage against the input distribution updated online in the current stage, based on the output feature differences. After this, the process ends.

[0087] Therefore, the information processing method according to the embodiments of this disclosure can utilize the differences in output features of the feature generation unit for the same input distribution to train the feature generation unit, thereby effectively reducing the forgetting problem in OUDA.

[0088] The following is combined with Figure 9 To describe another embodiment of the information processing method according to this disclosure.

[0089] like Figure 9 As shown, the information processing method according to an embodiment of this disclosure begins at step S210. In step S210, a loss function is generated to represent the output feature difference. Here, the output feature difference is the difference in output features between a feature generation unit trained in a predetermined stage and a feature generation unit trained in the current stage for the same specific input distribution.

[0090] Specifically, the same specific input distribution belongs to the same labeled source domain dataset. The feature generation unit trained in the predetermined stage is the feature generation unit trained with only the source domain dataset as the input distribution, and the feature generation unit trained in the current stage is the feature generation unit trained with the input distribution based on the source domain dataset, and the target domain dataset is added in stages, where the current stage is any one of multiple stages of adding the target domain dataset.

[0091] As an example, the loss function is the loss function of the output features of the feature generation unit in the current stage for the source domain dataset and the output features of the feature generation unit for the source domain dataset when the input distribution only contains the source domain dataset.

[0092] As another example, the loss function is the sum of the loss functions of the output features of the feature generation unit in the current stage for the source domain dataset and the output features of the feature generation units in all previous stages for the source domain dataset.

[0093] Furthermore, the loss function mentioned above can be any distance loss function, such as L2 distance loss function, L1 distance loss function, cosine distance loss function, etc.

[0094] Next, in step S220, the objects in the input distribution of the current stage are classified based on the output features of the feature generation unit of the current stage for the input distribution of the current stage.

[0095] Next, in step S230, the feature generation unit of the current stage classifies whether the input distribution of the current stage belongs to the source domain or the target domain based on the output features of the input distribution of the current stage.

[0096] Next, in step S240, the feature generation unit for the current stage is trained using the classification results based on the loss function. After this, the process ends.

[0097] Therefore, the information processing method according to the embodiments of this disclosure can minimize the difference in output features while preserving the model's learning ability on the target domain dataset by using a loss function based on source domain feature distillation, thereby solving the forgetting problem of OUDA and effectively improving the generalization and convergence of the model in online learning.

[0098] Various specific implementations of the above steps of the information processing method according to the embodiments of this disclosure have been described in detail above and will not be repeated here.

[0099] Obviously, the various operational processes of the information processing method according to this disclosure can be implemented as a computer executable program stored in various machine-readable storage media.

[0100] Furthermore, the objective of this disclosure can also be achieved by providing a storage medium storing the aforementioned executable program code directly or indirectly to a system or device, and having a computer or central processing unit (CPU) in the system or device read and execute the aforementioned program code. In this case, as long as the system or device has the function of executing a program, the implementation of this disclosure is not limited to a program, and the program can be in any form, such as an object program, a program executed by an interpreter, or a script program provided to an operating system.

[0101] The aforementioned machine-readable storage media include, but are not limited to: various memories and storage units, semiconductor devices, disk units such as optical, magnetic and magneto-optical disks, and other media suitable for storing information.

[0102] Alternatively, the technical solution of this disclosure can also be implemented by connecting to a corresponding website on the Internet, downloading and installing the computer program code according to this disclosure onto the computer, and then executing the program.

[0103] Figure 10 This is a block diagram illustrating an exemplary structure of a general-purpose personal computer in which information processing apparatus and methods according to embodiments of the present disclosure can be implemented.

[0104] like Figure 10 As shown, CPU 1301 executes various processes based on programs stored in read-only memory (ROM) 1302 or programs loaded into random access memory (RAM) 1303 from storage section 1308. RAM 1303 also stores data required as needed when CPU 1301 executes various processes, etc. CPU 1301, ROM 1302, and RAM 1303 are interconnected via bus 1304. Input / output interface 1305 is also connected to bus 1304.

[0105] The following components are connected to the input / output interface 1305: input section 1306 (including keyboard, mouse, etc.), output section 1307 (including display, such as cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.), storage section 1308 (including hard disk, etc.), and communication section 1309 (including network interface card, such as LAN card, modem, etc.). The communication section 1309 performs communication processing via a network, such as the Internet. Drive 1310 may also be connected to the input / output interface 1305 as needed. Removable media 1311, such as disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1310 as needed, so that computer programs read from them can be installed into storage section 1308 as needed.

[0106] When the above series of processes are implemented by software, the program constituting the software is installed from a network such as the Internet or a storage medium such as removable media 1311.

[0107] Those skilled in the art will understand that such storage media are not limited to Figure 10 The illustrated removable medium 1311 stores a program and is distributed separately from the device to provide the program to the user. Examples of removable media 1311 include magnetic disks (including floppy disks (registered trademark)), optical disks (including optical disc read-only memory (CD-ROM) and digital versatile disks (DVD)), magneto-optical disks (including mini-disk (MD) (registered trademark)), and semiconductor memory. Alternatively, the storage medium may be ROM 1302, a hard disk included in storage section 1308, etc., containing programs and distributed to the user along with the device containing them.

[0108] In the systems and methods of this disclosure, it is apparent that the components or steps can be decomposed and / or recombined. Such decomposition and / or recombination should be considered equivalent to those disclosed. Furthermore, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order. Some steps can be performed in parallel or independently of each other.

[0109] While embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings, it should be understood that the embodiments described above are merely illustrative and do not constitute a limitation thereof. Those skilled in the art can make various modifications and alterations to the above embodiments without departing from the spirit and scope of the present disclosure. Therefore, the scope of the present disclosure is defined only by the appended claims and their equivalents.

[0110] Regarding the implementation methods including the above embodiments, the following notes are also disclosed:

[0111] Appendix 1. An information processing apparatus for online unsupervised domain adaptation, comprising:

[0112] Feature generation unit, configured to generate features for the input distribution;

[0113] A difference generation unit, configured to generate output feature differences between a feature generation unit trained in a predetermined stage and a feature generation unit trained in the current stage, given the same specific input distribution; and

[0114] An update unit is configured to train the current stage's feature generation unit based on the output feature differences using the output features of the currently trained feature generation unit against the input distribution updated online for the current stage.

[0115] Appendix 2. The information processing apparatus according to Appendix 1, wherein,

[0116] The same specific input distribution refers to the same labeled source domain dataset.

[0117] Appendix 3. The information processing apparatus according to Appendix 2, wherein,

[0118] The feature generation unit trained in the predetermined phase is the feature generation unit trained with the input distribution consisting only of the source domain dataset, and

[0119] The feature generation unit trained in the current stage is a feature generation unit trained in stages with the input distribution added to the source domain dataset in stages, wherein the current stage is any one of multiple stages of adding the target domain dataset.

[0120] Appendix 4. The information processing apparatus according to Appendix 2 further includes:

[0121] A loss function generation unit is configured to generate a loss function to represent the differences in the output features.

[0122] Appendix 5. The information processing apparatus according to Appendix 4, wherein,

[0123] The loss function is the loss function of the output features of the feature generation unit in the current stage for the source domain dataset and the loss function of the output features of the feature generation unit for the source domain dataset when the input distribution only contains the source domain dataset.

[0124] Appendix 6. The information processing apparatus according to Appendix 5, wherein,

[0125] The loss function is a distance loss function.

[0126] Appendix 7. The information processing apparatus according to Appendix 4, wherein,

[0127] The loss function is the sum of the loss functions of the output features of the feature generation unit in the current stage for the source domain dataset and the output features of the feature generation units in previous stages for the source domain dataset.

[0128] Appendix 8. The information processing apparatus according to any one of Appendices 4 to 7, wherein,

[0129] The update unit includes a training unit, which is configured to train the feature generation unit of the current stage based on the loss function using the analysis results of the output features of the feature generation unit of the current stage for the input distribution of the current stage, so as to update the feature generation unit in real time.

[0130] Appendix 9. The information processing apparatus according to Appendix 8, wherein,

[0131] The updating unit further includes a classification unit, which classifies objects in the input distribution of the current stage based on the output features of the feature generation unit for the input distribution of the current stage, and provides the classification results to the training unit for training the feature generation unit.

[0132] Note 10. The information processing apparatus according to Note 9, wherein,

[0133] The updating unit further includes a domain discrimination unit, which classifies whether the input distribution of the current stage belongs to the source domain or the target domain based on the output features of the current stage feature generation unit for the current stage input distribution, and provides the classification result to the training unit for training the feature generation unit.

[0134] Note 11. The information processing apparatus according to Note 6, wherein,

[0135] The distance loss function is one of the L2 distance loss function, L1 distance loss function, and cosine distance loss function.

[0136] Appendix 12. A method for adaptive information processing in online unsupervised domains, comprising:

[0137] Generate the difference in output features between a feature generation unit trained in a predetermined stage and a feature generation unit trained in the current stage, given the same specific input distribution; and

[0138] Based on the output feature differences, the feature generation unit trained in the current stage is used to train the feature generation unit of the current stage with the output features of the input distribution updated online in the current stage.

[0139] Appendix 13. According to the information processing method described in Appendix 12, wherein,

[0140] The same specific input distribution refers to the same labeled source domain dataset.

[0141] Appendix 14. According to the information processing method described in Appendix 13, wherein,

[0142] The feature generation unit trained in the predetermined phase is the feature generation unit trained with the input distribution consisting only of the source domain dataset, and

[0143] The feature generation unit trained in the current stage is a feature generation unit trained in stages with the input distribution added to the source domain dataset in stages, wherein the current stage is any one of multiple stages of adding the target domain dataset.

[0144] Appendix 15. The information processing method according to Appendix 13 further includes:

[0145] Generate a loss function to represent the differences in the output features.

[0146] Appendix 16. According to the information processing method described in Appendix 15, wherein,

[0147] The loss function is the loss function of the output features of the feature generation unit in the current stage for the source domain dataset and the loss function of the output features of the feature generation unit for the source domain dataset when the input distribution only contains the source domain dataset.

[0148] Appendix 17. According to the information processing method described in Appendix 16, wherein,

[0149] The loss function is a distance loss function.

[0150] Appendix 18. According to the information processing method described in Appendix 15, wherein,

[0151] The loss function is the sum of the loss functions of the output features of the feature generation unit in the current stage for the source domain dataset and the output features of the feature generation units in previous stages for the source domain dataset.

[0152] Note 19. The information processing method according to any one of Notes 15 to 18, wherein,

[0153] Training the feature generation unit for the current stage includes: using the analysis results of the output features of the feature generation unit for the current stage based on the loss function to train the feature generation unit for the current stage input distribution, so as to update the feature generation unit in real time.

[0154] Appendix 20. A machine-readable storage medium carrying a program product including machine-readable instruction code stored thereon, wherein, when read and executed by a computer, the instruction code enables the computer to perform the information processing method according to Appendices 12-19.

Claims

1. An information processing apparatus for online unsupervised domain adaptation, comprising: a feature generation unit configured to generate features for an input distribution, wherein the input distribution is images; a difference generation unit configured to generate an output feature difference of a feature generation unit trained at a predetermined stage and a feature generation unit trained at a current stage for the same specific input distribution; and an update unit configured to train the feature generation unit at the current stage based on output features of the input distribution at the current stage updated online using the feature generation unit trained at the current stage based on the output feature difference, wherein the same specific input distribution is the same source domain dataset with labels, wherein the feature generation unit trained at the predetermined stage is a feature generation unit trained with only the source domain dataset as the input distribution, and the feature generation unit trained at the current stage is a feature generation unit trained with the input distribution increased by a target domain dataset without labels on a stage-by-stage basis based on the source domain dataset.

2. The information processing apparatus according to claim 1, wherein The current stage is any one of a plurality of stages in which the target domain dataset is increased. 3.The information processing apparatus according to claim 1, further comprising: a loss function generation unit configured to generate a loss function for representing the output feature difference. 4.The information processing apparatus according to claim 3, wherein the loss function is a loss function of output features of the feature generation unit at the current stage for the source domain dataset and output features of the feature generation unit trained with only the source domain dataset as the input distribution for the source domain dataset. 5.The information processing apparatus according to claim 4, wherein the loss function is a distance loss function. 6.The information processing apparatus according to claim 3, wherein the loss function is a sum of loss functions of output features of the feature generation unit at the current stage for the source domain dataset and output features of the feature generation units at respective stages before the current stage for the source domain dataset. 7.The information processing apparatus according to any one of claims 3 to 6, wherein the update unit includes a training unit configured to train the feature generation unit at the current stage based on an analysis result of output features of the feature generation unit at the current stage for the input distribution at the current stage using the loss function to update the feature generation unit in real time. 8.An information processing method for online unsupervised domain adaptation, comprising: generating an output feature difference of a feature generation unit trained at a predetermined stage and a feature generation unit trained at a current stage for the same specific input distribution; and training the feature generation unit at the current stage based on output features of the input distribution at the current stage updated online using the feature generation unit trained at the current stage based on the output feature difference, wherein the input distribution is images, wherein the same specific input distribution is the same source domain dataset with labels, wherein the feature generation unit trained at the predetermined stage is a feature generation unit trained with only the source domain dataset as the input distribution, and the feature generation unit trained at the current stage is a feature generation unit trained with the input distribution increased by a target domain dataset without labels on a stage-by-stage basis based on the source domain dataset. Among them, the feature generation unit trained in the predetermined stage is a feature generation unit trained in the case where the input distribution is only the source domain dataset, and the feature generation unit trained in the current stage is a feature generation unit trained in the case where the input distribution is gradually increased on the basis of the source domain dataset by the target domain dataset without labels.

9. A machine-readable storage medium having carried thereon a program product comprising machine-readable instruction code stored therein, wherein, The instruction code, when read and executed by a computer, can enable the computer to execute the information processing method according to claim 8. The instruction code, when read and executed by a computer, can enable the computer to execute the information processing method according to claim 8.

Images