Training device

Abstract

This training device receives input of a first dataset of labeled image data. Thereafter, the training device creates one or more second datasets (pseudo domains) obtained by performing, on the first dataset, at least one of image conversion for changing the painting style and image conversion for data expansion of the inputted image data. Next, the training device performs training by an MDG algorithm on a machine learning model that classifies the image data by using the first dataset and a pseudo multi-domain including one or more pseudo-domains.

Description

Learning device

[0001] The present invention relates to a learning device for a machine learning model.

[0002] The Deep Neural Network (DNN) used in image classification and the like may significantly deteriorate in classification accuracy under the distribution shift of input data generated by various factors such as weather and noise. For example, as shown in FIG. 1, when noise is added to the input data (image data) to the Classification Network, it may be impossible to correctly classify the input data.

[0003] For example, even if a distribution shift occurs in the input data to the DNN model used in applications such as autonomous driving, where the environment is likely to change and high safety is required, it is required to maintain the classification accuracy.

[0004] Here, as a technique for maintaining the classification accuracy for input data having a distribution different from the distribution of the data used during training of the DNN model, there is Multi-source Domain Generalization (MDG). MDG is a technique for learning a DNN model using a data set (multi-domain data set) composed of a plurality of domains. Many studies have been conducted on this MDG.

[0005] Gulrajani, Ishaan, and David Lopez-Paz. "In Search of Lost Domain Generalization." International Conference on Learning Representations., [online], [searched on October 31, 2024], Internet <URL: https: / / arxiv.org / abs / 2007.01434>

[0006] However, the creation of the multi-domain datasets required for MDG is costly, making it impractical to use MDG for training DNN models. For example, it is unclear which domains' data should be collected to create the multi-domain data used for MDG. Furthermore, the cost of labeling the data for each domain is also high.

[0007] Therefore, the present invention aims to solve the aforementioned problems and to realize low-cost training of machine learning models using the MDG algorithm.

[0008] To solve the aforementioned problems, the present invention is characterized by comprising: an input receiving unit that receives input of a first dataset of labeled image data; a dataset creation unit that creates one or more second datasets by performing at least one of image transformations that change the style of the input image data and image transformations that extend the data on the input first dataset; and a learning unit that performs training on a machine learning model that classifies the input image data using the input first dataset and one or more of the second datasets, using the MDG (Multi-source Domain Generalization) algorithm.

[0009] According to the present invention, it is possible to train machine learning models using the MDG algorithm at a low cost.

[0010] Figure 1 illustrates the decrease in classification accuracy due to distribution shift. Figure 2 shows an example of image transformation. Figure 3 illustrates an overview of the learning device. Figure 4 shows an example of the configuration of the learning device. Figure 5 shows an example of the processing procedure performed by the learning device. Figure 6 illustrates the dataset used in the experiment. Figure 7 illustrates the experimental results (1) of a pseudo-multidomain study using data augmentation. Figure 8 illustrates the experimental results (2) of a pseudo-multidomain study using data augmentation. Figure 9 illustrates the experimental results (3) of a pseudo-multidomain study using data augmentation. Figure 10 illustrates the experimental results (1) of a pseudo-multidomain study using style transfer. Figure 11 illustrates the experimental results (2) of a pseudo-multidomain study using style transfer. Figure 12 illustrates the experimental results (3) of a pseudo-multidomain study using style transfer. Figure 13 shows an example of a computer that runs the learning program.

[0011] The following describes embodiments for carrying out the present invention with reference to the drawings. The present invention is not limited to these embodiments.

[0012] [Overview] The overview of the learning device of this embodiment will be described with reference to Figure 2. For example, the learning device expands a single-domain dataset into a multi-domain dataset. The learning device then applies the MDG algorithm to the expanded multi-domain dataset (a pseudo-multi-domain dataset) to train a machine learning model (for example, a DNN (Deep Neural Network) model).

[0013] Here, the learning device performs the following image transformation to expand the single-domain dataset into a multi-domain dataset.

[0014] (1) Data augmentation: Data augmentation is an image transformation used to improve the diversity of training data for DNN models.

[0015] Examples: MixUp [ICLR18], CutMix [ICCV19], AugMix [ICLR20], IPMix [NeurIPS23] (see Figure 2), RandAugment [CVPRW20], TrivialAugment [ICCV21], RandConv [ICLR21]

[0016] IPMix is ​​described in the following reference 1: Reference 1: Huang, Zhenglin, et al. "IPMix: label-preserving data augmentation method for training robust classifiers." Advances in Neural Information Processing Systems 36 (2023): 63660-63673.

[0017] (2) Style Transfer Style transfer is an image transformation that changes the style of an image using a DNN (for example, a GAN (Generative Adversarial Network)).

[0018] Examples: Style Transfer [ICCV17], Edge Detection [WACV20], Cartoon GAN [CVPR18] (See Figure 2 for the above)

[0019] The learning device applies the above image transformation to a single-domain dataset to create a pseudo-multi-domain dataset (hereinafter referred to as "pseudo-multi-domain" as appropriate). For example, as shown in Figure 3, the learning device generates n transformation candidates ({O0, O1, ... O) based on the expected distribution shift. n Select N transformations from the following (multiple selections are allowed).

[0020] The learning device then applies the selected N transformations to a single-domain dataset to create a pseudo-multi-domain dataset.

[0021] For example, if the selected transformation is {No Transformation, IPMix, IPMix, Style Transfer}, the learning device applies the above transformation to a single-domain dataset to create Pseudo-Domain0, Pseudo-Domain1, Pseudo-Domain2, and Pseudo-Domain3. In other words, the learning device creates a pseudo-multi-domain consisting of the original domain dataset, two pseudo-domains created by applying IPMix to the original domain dataset, and one pseudo-domain created by applying Style Transfer to the original domain dataset.

[0022] Subsequently, the learning device trains the DNN model using the MDG algorithm with the created pseudo-multi-domain model. This allows the learning device to train a DNN model using MDG at a low cost.

[0023] In the following explanation, we will use the example of a DNN model as the machine learning model that the learning device is training on, but the explanation is not limited to this.

[0024] [Configuration Example] Next, an example of the configuration of the learning device 10 will be described using Figure 4. The learning device 10 includes, for example, an input / output unit 11, a storage unit 12, and a control unit 13.

[0025] The input / output unit 11 is an interface that handles the input and output of various types of data. For example, the input / output unit 11 accepts input of a single-domain dataset. Also, for example, the input / output unit 11 accepts input for selecting image transformations used to create a pseudo-multi-domain dataset.

[0026] The memory unit 12 stores data, programs, etc., that are referenced when the control unit 13 performs various processes. The memory unit 12 is implemented by semiconductor memory elements such as RAM (Random Access Memory) and flash memory, or by storage devices such as hard disks and optical discs. For example, the memory unit 12 stores the single-domain dataset, parameters of the DNN model to be trained, and conversion candidate information. The DNN model is, for example, a DNN model that has been previously trained on ImageNet or the like.

[0027] The conversion candidate information indicates candidate image conversions to be used in creating pseudo-multidomains (for example, image conversions for data augmentation, image conversions for style transformation, etc.). This conversion candidate information is input in advance via the input / output unit 11. The storage unit 12 stores the pseudo-multidomains created by the control unit 13.

[0028] The control unit 13 is responsible for controlling the entire learning device 10. The functions of the control unit 13 are realized, for example, by the CPU (Central Processing Unit) executing a program stored in the memory unit 12.

[0029] The control unit 13 includes an input receiving unit 131, an image conversion unit (dataset creation unit) 132, and a learning unit 133. The selection unit 134 and classification unit 135, shown by dashed lines, may or may not be included; the cases in which they are included will be described later.

[0030] The input receiving unit 131 accepts input of a dataset (first dataset) to be used to create a pseudo-multi-domain dataset. For example, the input receiving unit 131 accepts input of a dataset of single-domain labeled image data.

[0031] The image conversion unit 132 creates a pseudo-domain (second dataset) from the input dataset using two types of image conversion (image conversion for changing the image style and image conversion for expanding the data). The image conversion unit 132 then outputs the input dataset and one or more pseudo-domains created using this dataset as a pseudo-multi-domain.

[0032] For example, the image conversion unit 132 creates a pseudo-domain by applying an image conversion selected by the user from among the image conversions shown in the conversion candidate information to the input dataset. The image conversion unit 132 then outputs the input dataset and the created pseudo-domain group as a pseudo-multi-domain. Note that the pseudo-multi-domain may contain multiple pseudo-domains to which the same image conversion has been applied.

[0033] For example, consider a case where the image conversion unit 132 receives a selection from the user for {no conversion, IPMix, IPMix, Style Transfer} as the image conversion to be used to create a pseudo-multi-domain. In this case, the image conversion unit 132 performs the above {no conversion, IPMix, IPMix, Style Transfer} image conversion on the input single-domain dataset to create a pseudo-multi-domain that includes the original dataset, two pseudo-domains created by image conversion of the original dataset using IPMix, and one pseudo-domain created by image conversion of the original dataset using Style Transfer.

[0034] The image conversion unit 132 may create a pseudo-domain by applying multiple types of image conversions to the original dataset. For example, the image conversion unit 132 may create a pseudo-domain by applying image conversion by IPMix and image conversion by Style Transfer to the original dataset.

[0035] The learning unit 133 performs training on the DNN model to be trained using the MDG algorithm, with respect to the pseudo-multi-domain created by the image conversion unit 132.

[0036] With this type of learning device 10, it is possible to train a DNN model using the MDG algorithm at a low cost.

[0037] [Example of Processing Procedure] Next, an example of a processing procedure performed by the learning device 10 will be explained using Figure 5. First, the input receiving unit 131 of the learning device 10 receives input of a dataset to be used to create a pseudo-multi-domain (S1). The learning device 10 also receives input for selecting image transformations to be used to create a pseudo-multi-domain (S2). For example, the learning device 10 displays transformation candidate information to the user and receives input from the user for selecting image transformations to be used to create a pseudo-multi-domain.

[0038] For example, if preliminary testing determines that there is a possibility of a distribution shift due to style changes in the image data input to the trained DNN model, the user will select to include style transformation in the image transformation used to create the pseudo-multidomain model. On the other hand, if it is determined that there is no possibility of a distribution shift due to style changes in the image data input to the trained DNN model, the user will not include style transformation as an image transformation (only data augmentation will be performed).

[0039] This is because the inventors' experimental results (details below) confirmed that when there is a possibility of a distribution shift due to changes in painting style in the image data input to the trained DNN model, using style transformation to create pseudo-multidomains improves the classification accuracy of the DNN model trained using the MDG algorithm.

[0040] After S2, the image transformation unit 132 performs the image transformation selected in S2 on the dataset input in S1 to create a pseudo-multidomain (S3). Subsequently, the learning unit 133 uses the pseudo-multidomain created in S3 to train a DNN model using the MDG algorithm (S4).

[0041] By having the learning device 10 perform the above processing, it is possible to train a DNN model using the MDG algorithm at a low cost.

[0042] [Other Embodiments] The learning device 10 may further include, for example, a selection unit 134 shown in FIG. 4. When the selection unit 134 determines, through a preliminary test, that there is a possibility of a distribution shift in the画风 change in the image data input to the DNN model after learning, it instructs the image conversion unit 132 to use image conversion with a画风 change for creating a pseudo multi-domain.

[0043] For example, when the selection unit 134 performs a preliminary test on the image data input to the DNN model after learning and determines that there is a possibility of a distribution shift in the画风 change in the image data input to the DNN model, it instructs the image conversion unit 132 to use image conversion with a画风 change (e.g., Style Transfer) for creating a pseudo multi-domain.

[0044] Also, when the selection unit 134 determines, through the above preliminary test, that there is no possibility of a distribution shift in the画风 change in the image data input to the DNN model after learning, it instructs the image conversion unit 132 not to use image conversion with a画风 change for creating a pseudo multi-domain (e.g., apply only image conversion for data augmentation such as IPMix).

[0045] Then, the image conversion unit 132 applies the image conversion instructed by the selection unit 134 to the input data set to create a pseudo multi-domain. After that, the learning unit 133 learns the DNN model using the created pseudo multi-domain by the MDG algorithm.

[0046] As a result, the learning unit 133 can learn the DNN model by the MDG algorithm according to the assumed distribution shift in the image data input to the DNN model after learning. As a result, the learning unit 133 can construct a DNN model with high robustness against distribution shifts.

[0047] Note that the learning device 10 may further include a classification unit 135 (see FIG. 4) that classifies input data using the DNN model after learning.

[0048] [Experimental Results] Next, we will describe the evaluation experiment of the DNN model trained by the learning device 10 of this embodiment. The datasets used in the evaluation experiment are the four datasets shown in Figure 6 (VLCS, PACS, Office-Home, Terra Incognita). Each dataset contains four domains. The DNN model used in the experiment is a DNN model that has been previously trained on ImageNet, etc.

[0049] In each experiment, data from one domain of the dataset was used to train the DNN model, and the remaining three domains were used to evaluate the classification accuracy of the trained DNN model.

[0050] [Experimental Results of Pseudo-Multi-Domain Modeling Using Data Augmentation (1)] This section describes the experimental results of pseudo-multi-domain modeling using data augmentation. The pseudo-multi-domain model in this experiment consists of the original (single-domain dataset) and a dataset created by applying image transformations using data augmentation to the original dataset. The classification accuracy was evaluated using the difference (Gain) between the classification accuracy when trained using ERM (Empirical Risk Minimization Algorithm).

[0051] Figure 7 shows the experimental results. As shown in Figure 7, the gains were high when using EQRM (Experiential Quantile Risk Minimization), SD, and SelfReg among the MDG algorithms. Furthermore, it was confirmed that the gains were high when applying data augmentation methods such as AugMix and IPMix.

[0052] [Experimental Results of Pseudo-Multi-Domain Data Using Data Augmentation (2)] Next, we confirmed whether algorithms effective in MDG are also effective for training DNN models in the learning device 10. The classification accuracy of the DNN models used for comparison was the classification accuracy when a DNN model trained on three domains of the dataset using each MDG algorithm was tested on the remaining domain. The pseudo-multi-domain dataset used by the learning device 10 consists of the original (single-domain dataset) and a dataset to which IPMix has been applied.

[0053] The experimental results are shown in Figure 8. In Figure 8, the vertical axis represents the accuracy of the DNN model trained by the learning device 10, and the horizontal axis represents the classification accuracy of the DNN model used for comparison. As shown in Figure 8, there is a correlation between the classification accuracy of the two, confirming that the algorithm effective in MDG is also effective in the learning device 10.

[0054] [Experimental Results of Pseudo-Multidomain Using Data Augmentation (3)] Next, an evaluation experiment was conducted on combinations of data augmentation types used to create pseudo-multidomains. The MDG algorithm used in the experiment was SD. The experimental results are shown in Figure 9.

[0055] As shown in Figure 9, it was confirmed that using IPMix to create pseudo-multi-domains best improved the classification accuracy of the DNN model. Furthermore, the classification accuracy was IPMix x 2 > IPMix x 3 > IPMix x 1. In addition, it was found that increasing the number of pseudo-domains contributes to improved classification accuracy depending on the dataset.

[0056] Furthermore, it was confirmed that using multiple IPMix combinations was effective in improving classification accuracy, as combinations of different data augmentations other than IPMix did not improve classification accuracy.

[0057] [Summary of Data Augmentation] From the experimental results above, it was confirmed that the learning device 10 can improve the classification accuracy of the DNN model compared to ERM by creating a pseudo-multi-domain using data augmentation and training the DNN model using the MDG algorithm with this pseudo-multi-domain.

[0058] Furthermore, regardless of the type of data augmentation used to create the pseudo-multidomain model, it was confirmed that using MDG algorithms such as EQRM, SD, and SelfReg to train the DNN model improved the classification accuracy of the DNN model compared to ERM.

[0059] Furthermore, it was confirmed that data augmentation techniques designed to improve robustness, such as IPMix, are effective for creating pseudo-multi-domains. Additionally, it was confirmed that using IPMix multiple times is more effective than combining different types of data augmentation techniques for creating pseudo-multi-domains.

[0060] [Experimental Results of Pseudo-Domains Using Style Transfer (1)] Next, we will explain the experimental results of pseudo-multi-domains using style transfer. The dataset used in the experiment was PACS (see Figure 6). PACS's Photo was used to create the pseudo-multi-domains. PACS's Art, Cartoon, and Sketch were used to evaluate the classification accuracy of the DNN model.

[0061] Figure 10 shows the experimental results. In Figure 10, P=Photo, A=Art, C=Cartoon, and S=Sketch.

[0062] In the experimental results below, the left side of the "→" indicates the data used to train the DNN model, and the right side of the "→" indicates the data used to evaluate the classification accuracy of the trained DNN model. For example, P→A indicates that the DNN model was trained using Photo, and the classification accuracy of the trained DNN model was evaluated using Art.

[0063] Furthermore, in this experiment, the correspondence between each domain of PACS and style transfer was considered to be Art = Style Transfer, Cartoon = Cartoon GAN, and Sketch = Edge Detection. Therefore, for example, if Style Transfer is used as the training pseudo-domain of the DNN model, it is obvious and unfair that high accuracy will be obtained in the Art domain. For this reason, in this experiment, we also conducted experiments in which Style Transfer was not included in the training pseudo-domain when evaluating accuracy in the Art domain. The † in Figure 10 shows the evaluation results when style transfer corresponding to the data being evaluated was not included in the pseudo-multi-domain, as described above. Note that in the experimental results described below, evaluation results where Pseudo-domain is blank show the evaluation results when a single domain was used to train the DNN model.

[0064] As shown in Figure 10, in the case of PACS, the classification accuracy was 61.37 when a pseudo-multi-domain dataset was used and the DNN model was trained using SD, 57.73 when a pseudo-multi-domain dataset was used and the DNN model was trained using ERM, and 54.30 when a single-domain dataset transformed with IPMix was trained using ERM. Therefore, in the case of PACS, it was confirmed that using style transfer to create pseudo-multi-domain datasets resulted in higher classification accuracy for the DNN model trained using the MDG algorithm compared to using data augmentation.

[0065] [Experimental Results of Pseudo-Domains Using Style Transfer (2)] Next, we will explain the experimental results when using Office-Home (see Figure 6). Office-Home's Real World was used to create pseudo-multi-domains. Office-Home's Art, Clipart, and Product were used to evaluate the classification accuracy of the DNN model.

[0066] Figure 11 shows the experimental results. In Figure 11, R = Real World, A = Art, C = Clipart, and P = Product. Note that the † in Figure 11 indicates the evaluation results when the style transformation corresponding to the data being evaluated was not included in the pseudo-multi-domain, as described above.

[0067] As shown in Figure 11, for Office-Home, the classification accuracy was 63.50 when training a DNN model using ERM with pseudo-multi-domain (Original+Style+Cartoon+Edge), 63.20 when training a DNN model using SD with pseudo-multi-domain (Original+Style+Cartoon+Edge), and 62.17 when training a DNN model using ERM with a single-domain dataset transformed by IPMix.

[0068] Therefore, in the case of Office-Home, it was confirmed that using style transfer to create pseudo-multi-domain models resulted in higher classification accuracy for the trained DNN model compared to using data augmentation.

[0069] [Experimental Results of Pseudo-Domains Using Style Transfer (3)] Next, we will explain the experimental results when using VLCS (see Figure 6). In this experiment, style transfer was performed on all domain datasets in VLCS to create pseudo-multi-domains. The classification accuracy was evaluated using the datasets of domains that were not used to create the pseudo-multi-domains.

[0070] Figure 12 shows the experimental results. In Figure 12, C = Caltech101, L = LabelMe, S = SUN09, and V = VOC2007.

[0071] As shown in Figure 12, in the case of VLCS, the classification accuracy when training a DNN model using ERM with a single-domain dataset transformed by IPMix was 64.63, the classification accuracy when training a DNN model using SD with pseudo-multi-domain (Original + Style + Cartoon + Edge) was 61.33, and the classification accuracy when training a DNN model using ERM with pseudo-multi-domain (Original + Style + Cartoon + Edge) was 60.22.

[0072] Therefore, in the case of VLCS, it was confirmed that using style transfer to create pseudo-multi-domain models results in lower classification accuracy for the DNN model compared to training the DNN model on a single domain or using data augmentation to create pseudo-multi-domain models.

[0073] [Summary regarding style transfer] From the experimental results above, it was confirmed that applying style transfer to the dataset when the learning device 10 creates a pseudo-multi-domain dataset allows for accurate classification of data from domains composed of different styles, such as PACS and Office-Home. This is presumed to be because the learning device 10 incorporates differences in style into the training of its DNN model, thereby learning common features between styles.

[0074] Therefore, in the learning device 10, when it is predicted that a distribution shift due to style change will occur in the data input to the trained DNN model, it was confirmed that training the DNN model in a pseudo-multi-domain including style transformation, as described above, is effective.

[0075] Furthermore, it was confirmed that applying style transfer to the dataset when the learning device 10 creates a pseudo-multi-domain dataset is not effective for VLCS. This is presumed to be because VLCS is a domain composed of data collected from different cameras, and therefore, incorporating style differences into the training of the DNN model does not contribute to improving the classification accuracy of the DNN model.

[0076] [System Configuration, etc.] Furthermore, the components of each part shown in the diagram are functional concepts and do not necessarily need to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown in the diagram, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions. In addition, all or any part of the processing functions performed by each device can be realized by a CPU and the program executed on that CPU, or by hardware using wired logic.

[0077] Furthermore, among the processes described in the embodiments described above, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0078] [Program] The learning device 10 described above can be implemented by installing a program (learning program) as packaged software or online software on a desired computer. For example, by having the above program run on an information processing device, the information processing device can be made to function as the learning device 10. The information processing device referred to here includes mobile communication terminals such as smartphones, mobile phones and PHS (Personal Handyphone System), as well as terminals such as PDA (Personal Digital Assistant).

[0079] Figure 13 shows an example of a computer running a learning program. Computer 1000 has, for example, memory 1010 and a CPU 1020. Computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0080] Memory 1010 includes ROM (Read Only Memory) 1011 and RAM (Random Access Memory) 1012. ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, a display 1130.

[0081] The hard disk drive 1090 stores, for example, the OS 1091, application program 1092, program module 1093, and program data 1094. That is, the program that defines each process executed by the learning device 10 is implemented as a program module 1093 in which executable code for a computer is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, a program module 1093 for executing a process similar to the functional configuration of the learning device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0082] Furthermore, the data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, memory 1010 or hard disk drive 1090. The CPU 1020 then reads the program module 1093 and program data 1094 stored in memory 1010 or hard disk drive 1090 into RAM 1012 as needed and executes them.

[0083] Furthermore, the program module 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090; for example, they may be stored in a removable storage medium and read by the CPU 1020 via a disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via a network interface 1070.

[0084] 10 Learning device 11 Input / Output unit 12 Memory unit 13 Control unit 131 Input receiving unit 132 Image conversion unit 133 Learning unit 134 Selection unit 135 Classification unit

Claims

1. A learning device comprising: an input receiving unit that receives input of a first dataset of labeled image data; a dataset creation unit that creates one or more second datasets by performing at least one of image transformations that change the style of the input image data and image transformations that extend the data on the input first dataset; and a learning unit that performs training on a machine learning model that classifies input image data using the input first dataset and one or more of the second datasets, using the MDG (Multi-source Domain Generalization) algorithm.

2. The learning device according to claim 1, characterized in that the dataset creation unit creates the second dataset by applying IPMix to the input first dataset.

3. The learning device according to claim 1, further comprising a selection unit that, if prior testing determines that there is a possibility of a distribution shift in style changes occurring in the image data input to the machine learning model after training by the learning unit, instructs the dataset creation unit to create a second dataset by performing an image transformation that changes the style on the input first dataset.

4. The learning device according to claim 1, characterized in that the MDG algorithm is EQRM (Experiential Quantile Risk Minimization), SD, or SelfReg.

Images