Transfer learning device, transfer learning method, and transfer learning program

Abstract

This transfer learning device is provided with: a knowledge transfer learning unit (11) that, on the basis of a learned model relating to a first type of data, performs knowledge transfer learning using learning data relating to a second type of data different from the first type, and generates an intermediate model from the learned model; and a knowledge migration progress determination unit (12; and a determination unit (12, 12 ') that determines, in the generated intermediate model, whether or not the reference likelihood of the learning data to which the reference tag has been added is greater than the likelihood of each of a plurality of pieces of learning data to which a non-reference tag other than the reference tag has been added, by multiple comparative tests.

Description

Technical Field

[0001] The present disclosure relates to transfer learning techniques. Background Art

[0002] When processing data of a certain modality (type of data) for a new process, if there is a model asset that has been fully learned through machine learning for a modality different from the modality, sometimes you want to migrate and flexibly use the knowledge related to the model asset. Patent document 1 discloses a technology related to such knowledge migration. More specifically, Patent document 1 discloses the following method: learning is performed on learning data with labels attached, using the root mean square error (RMSE) as the loss function, setting the condition of terminating the learning without waiting for the maximum number of epochs of 50 if the loss does not decrease for 5 consecutive epochs as the end condition, using the mean absolute error (MAE) and the coefficient of determination (R 2 ) is used as an evaluation indicator to evaluate the accuracy of the destination modality of knowledge transfer.

[0003] Prior art literature

[0004] Patent Literature

[0005] Patent Document 1: Japanese Patent Application Laid-Open No. 2021-135726 Summary of the Invention

[0006] Problems to be solved by the invention

[0007] According to the method disclosed in Patent Document 1, an evaluation index for evaluating the accuracy of the knowledge transfer destination modality is used. Therefore, by setting an appropriate threshold value related to the evaluation index, the quality of knowledge transfer can be ensured.

[0008] However, according to the conventional method of searching for the minimum loss using a loss function as described in Patent Document 1, learning is performed equivalently regardless of whether the learning data is sufficiently learned data, and therefore there is a problem that a long learning time is required.

[0009] The present disclosure has been made to solve such a problem, and an object of the present disclosure is to provide a transfer learning technology that can shorten the learning time based on transfer learning.

[0010] Means for solving problems

[0011] One aspect of the transfer learning device of the embodiment of the present disclosure comprises: a knowledge transfer learning unit, which performs knowledge transfer learning based on a learned model related to data of a first category using learning data related to a second category different from the first category, and generates an intermediate model based on the learned model; and a knowledge transfer progress determination unit, which determines, in the generated intermediate model, by a multiple comparison test whether the baseline likelihood of the learning data to which a baseline label is attached is greater than the likelihood of a plurality of learning data to which non-baseline labels other than the baseline label are attached.

[0012] Effects of the Invention

[0013] According to the transfer learning device of the embodiment of the present disclosure, the learning time based on transfer learning can be shortened. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a diagram showing a configuration example of a transfer learning device according to the first embodiment.

[0015] Figure 2A This is a diagram showing a hardware configuration example of a transfer learning device.

[0016] Figure 2B This is a diagram showing a hardware configuration example of a transfer learning device.

[0017] Figure 3 This is a diagram showing the outline of transfer learning.

[0018] Figure 4A This is a likelihood transition graph showing the likelihood transition of the visible image based on the intermediate model generated by transfer learning.

[0019] Figure 4B This is a likelihood transition graph showing the likelihood transition of depth images based on the intermediate model generated by transfer learning.

[0020] Figure 5A This is a graph showing the target range of the p-value table created for the visual image.

[0021] Figure 5B This is a diagram showing the target range of the p-value table created for the depth image.

[0022] Figure 6A is a table of p values in S1 for visible and depth images.

[0023] Figure 6B is a table of p values in S2 for visible and depth images.

[0024] Figure 6C This is a table of p-values in S3 for visible and depth images.

[0025] Figure 7 This is a flowchart showing the transfer learning method of implementation mode 1.

[0026] Figure 8 This is a diagram showing a configuration example of a transfer learning device according to a second embodiment.

[0027] Figure 9A It is a p-value table about the depth image.

[0028] Figure 9B It is a p-value table about the depth image.

[0029] Figure 9C It is a p-value table about the depth image.

[0030] Figure 10 This is a flowchart showing the transfer learning method of implementation mode 2. DETAILED DESCRIPTION

[0031] Hereinafter, various embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. In the accompanying drawings, components labeled with the same or similar reference numerals have the same or similar structures or functions, and repeated descriptions of such components will be omitted.

[0032] Implementation method 1.

[0033] <Structure>

[0034] Reference Figures 1 to 6C , the transfer learning device 1 of embodiment 1 of the present disclosure is described. Figure 1 1 is a diagram showing the configuration of a transfer learning device 1 according to Embodiment 1. The transfer learning device 1 is a device that performs transfer learning to transfer knowledge related to a transfer source (source) to a transfer destination (target).

[0035] If reference Figure 3 To illustrate, transfer learning device 1, for example, transfers knowledge from model 1 related to a domain of visible images, which has been thoroughly trained on the nature of objects captured in visible images, to generate model 1' capable of determining the nature of objects captured in depth images. In other words, transfer learning device 1 performs heterogeneous transfer learning across different feature spaces but the same task.

[0036] exist Figure 3 In the stage "before knowledge transfer" shown in FIG, model 1 is a learned model that has fully learned visual images. Therefore, for a visual image VI that captures a certain object, it can appropriately determine what the object is. Figure 3In the , the likelihood of the correct label of the visible image is represented by gray. The likelihood of the correct label of the visible image is the largest, so it is appropriately determined (indicated by "0" in the figure). However, for the depth image DI that captures a certain object, model 1 cannot appropriately determine what the object is. Figure 3 In , the likelihood of the correct label of the depth image is indicated by shading. The likelihood of the correct label indicated by shading is not the maximum, and therefore, it is not properly determined (indicated by "×" in the figure).

[0037] exist Figure 3 In the stage of "after knowledge transfer" shown in FIG, model 1' is a learned model generated by transfer learning using depth images based on model 1. Therefore, model 1' can appropriately determine what the object is based on the depth image DI of a certain object. The likelihood of the correct label represented by the shadow is the largest, so it is appropriately determined (indicated by "0" in the figure). On the other hand, model 1' is trained on depth images, such as Figure 3 As shown, for a visual image VI that captures an object, the object cannot be properly determined. The likelihood of the correct label for the visual image is not maximized, so an appropriate determination is not made (indicated by an "×" in the figure). Furthermore, the degree of similarity between the domains of the knowledge transfer source and the knowledge transfer destination varies with the performance of model 1' in determining the domain of the knowledge transfer source.

[0038] exist Figure 3 In the process of knowledge transfer, the stage of "knowledge transfer in progress" between before and after knowledge transfer indicates that the judgment result of the intermediate model during knowledge transfer changes.

[0039] The transfer learning device 1 may also generate a single model, for example, model 1 ′, based on a plurality of models 1 to 3 related to the domain of the knowledge transfer source.

[0040] In order to carry out the above transfer learning, Figure 1 As shown, the transfer learning device 1 includes a knowledge transfer learning unit 11, a knowledge transfer progress determination unit 12, a knowledge transfer repetition determination unit 13, and a knowledge transfer result output unit 14. The transfer learning device 1 is connected to a storage device 2 that stores data related to knowledge of the knowledge transfer source, a storage device 3 that stores data related to knowledge of the knowledge transfer destination, and a storage device 4 that stores data related to knowledge of the knowledge transfer destination after knowledge transfer, thereby forming a transfer learning system.

[0041] The data stored in the storage device 2 is data of a learned model that has been fully learned. For example, data of a learned model for a visual image with a label, which has been fully learned to determine whether an object appearing in the visual image is a person or a car, is stored. Examples of such learned models include ImageNet. The object of the learned model can also be an image other than a visual image. For example, it can also be an infrared image or a thermal image. In addition, the object of the learned model is not limited to an image. For example, the object of the learned model can also be a two-dimensionally represented measurement value or frequency distribution such as reflectivity, reflection intensity, or frequency measured using a sensor not shown.

[0042] The data stored in storage device 3 is used for learning the destination of knowledge transfer. For example, it stores learning data related to depth images of people, cars, buildings, etc. Each depth image is labeled to indicate the object being displayed. Depth images, also known as range images, are images with distance values as pixel values. The stored learning data can be of a different modality from the data stored in storage device 2 and can be images other than depth images, or two-dimensional representations of measurement values or frequency distributions other than images.

[0043] The data stored in the storage device 4 stores data of the model after knowledge transfer. For example, when a new model is formed using the knowledge of the learned model for visible images stored in the storage device 2 and the learned data for depth images stored in the storage device 3, the model after knowledge transfer is the learned model for depth images.

[0044] (Knowledge Transfer Learning Department)

[0045] The following describes in more detail the structure of the transfer learning device 1. The following description is based on the case where the data stored in the storage device 2 is learned model data for visible images, and the data stored in the storage device 3 is learned data for depth images.

[0046] The knowledge transfer learning unit 11 is a functional unit that performs knowledge transfer learning based on a learned model related to a first type of data using learning data related to a second type of data different from the first type, and generates an intermediate model based on the learned model. The knowledge transfer learning unit 11 performs such learning when it first acquires the learned model and learning data, or when it receives the output of repeated knowledge transfer learning from the knowledge transfer iteration determination unit 13. More specifically, the knowledge transfer learning unit 11 acquires data such as the structure of the neural network and node weight coefficient data of the learned model related to the visual image from the storage device 2 as data related to the modality (data type) of the knowledge transfer source. Furthermore, the knowledge transfer learning unit 11 acquires learning data related to depth images from the storage device 3 as data related to the modality of the knowledge transfer destination. Using this acquired data, the knowledge transfer learning unit 11 performs knowledge transfer learning based on the learned model related to the visual image using the learning data related to the depth image. Knowledge transfer learning is performed using publicly known techniques such as those described in Patent Document 1. The knowledge transfer learning unit 11 generates an intermediate model in the process of knowledge transfer by performing knowledge transfer learning, and outputs the generated intermediate model. The output intermediate model may be temporarily stored in a storage device (not shown).

[0047] Through transfer learning, the generated intermediate model gradually becomes suitable for the modality of the knowledge transfer destination. Figure 4A and Figure 4B 3 is a diagram showing the state of this deformation. Figure 4A This is a likelihood transition graph showing the likelihood transition of the visible image based on the intermediate model generated by transfer learning. Figure 4B This is a likelihood transition graph showing the likelihood transition of depth images based on the intermediate model generated by transfer learning. Assume that the correct label is label 1 in both the knowledge transfer source and the knowledge transfer destination.

[0048] exist Figure 4A In the example, the model used as the basis is a previously learned model for visible images. Therefore, in the stage immediately after transfer learning begins, for example, during epochs 1 to 10, the generated intermediate model determines with high likelihood that the input visible image is the image of label 1. However, as transfer learning progresses, the intermediate model gradually transforms into a model for depth images. Therefore, after a period of time has passed since the start of transfer learning, for example, during epochs 41 to 50, the generated intermediate model is no longer able to determine with high likelihood that the input visible image is the image of label 1.

[0049] As the opposite of this situation, Figure 4BAs shown, the intermediate model cannot determine with high likelihood that the input depth image is an image of label 1 at the beginning of transfer learning, but after a period of time after the start of transfer learning, it can determine with high likelihood that the input depth image is an image of label 1.

[0050] exist Figure 4A and Figure 4B In the example, it is assumed that the model used as the basis is a single learned model for visual images. However, multiple learned models for visual images can also be used as the basis. The multiple models used as the basis are usually learned from different data sets even for visual images with the same label. Therefore, by integrating and transferring the knowledge of multiple models, the judgment accuracy of the generated learning model can be improved. Assume that the models used as the basis are Model 1, Model 2, and Model 3. In this case, Figure 4A and Figure 4B In the process, round numbers 1, 4, 7, 10, ... are assigned to the rounds of model 1, round numbers 2, 5, 8, 11, ... are assigned to the rounds of model 2, and round numbers 3, 6, 9, 12, ... are assigned to the rounds of model 3. The three consecutive rounds (round numbers 1 to 3, 4 to 6, 7 to 9, 10 to 12, ...) are treated as a unit, thereby enabling the integrated transfer of knowledge from multiple learned models.

[0051] (Knowledge Transfer Progress Assessment Department)

[0052] The knowledge transfer progress determination unit 12 is a functional unit that determines the progress of knowledge transfer learning. More specifically, within the intermediate model generated by the knowledge transfer learning unit 11, the knowledge transfer progress determination unit 12 uses a multiple comparison test to determine whether the baseline likelihood of learning data with a baseline label is greater than the likelihoods of multiple learning data with non-baseline labels other than the baseline label. Because it is unclear what degree of likelihood difference is required to determine a significant relationship, the multiple comparison test is used to determine the progress of knowledge transfer learning based on common indicators such as confidence levels or confidence intervals.

[0053] A multiple comparison test is a method for testing which groups have differences when measuring values for three or more groups. As a method for performing a multiple comparison test, for example, Dunnett's multiple comparison test described in the following non-patent document 1 can be used. Furthermore, the method of the multiple comparison test is not limited to the Dunnett method, and other methods such as the Williams test and the Tukey test may also be used.

[0054] Non-Patent Literature 1: Takaaki Shiraishi and Hiroshi Sugiura, “Theory and Numerical Calculation of Multiple Comparison Method,” Kyoritsu Publishing Co., Ltd., pp. 76-81 (2018).

[0055] In Dunnett's multiple comparison test, when comparing the mean values of a measurement group, a value called a "tolerance" is calculated from multiple trial measurement groups based on the confidence level p. If the mean value of the reference sample and the mean values of each sample exceed this tolerance and a significant relationship exists, the significant relationship is considered to exist at the confidence level p. The value of p that defines the boundary between the significant relationship and the significant relationship is called the p-value.

[0056] The knowledge transfer progress determination unit 12 calculates a p-value based on the likelihood of the image data input to the intermediate model to determine, using a multiple comparison test, whether the baseline likelihood of the learning data with the baseline label is greater than the likelihoods of multiple learning data with non-baseline labels other than the baseline label. Because variations occur within each round, the p-value is calculated based on the likelihoods of the image data across multiple rounds.

[0057] Figures 6A to 6C This is a table showing examples of calculated p-values. Figure 6A shows the likelihood transition of a visual image. Figure 5A The p-value table of section S1 in the , and the likelihood transition of the depth image Figure 5B In addition, Figure 6B shows the likelihood transition of a visual image. Figure 5B The p-value table of segment S2 in the , and the likelihood transition representing the depth image Figure 5B In addition, Figure 6C The p-value table of the segment S3 in FIG. 5C showing the likelihood transition of the visible image and the p-value table showing the likelihood transition of the depth image are shown. Figure 5B In addition, Figure 5A The likelihood transfer and Figure 4A The likelihood transfer of is the same, Figure 5B The likelihood transfer and Figure 4BThe likelihood transition is the same. In addition, segment S1 is a segment of rounds 1 to 10, segment S2 is a segment of rounds 21 to 30, and segment S3 is a segment of rounds 41 to 50. Through a single multiple comparison test, the size relationship between a certain label (for example, label 1) and multiple other labels (for example, labels 2 to 6) can be evaluated based on a certain label (for example, label 1), and the p-value of each cell in a column of the p-value table (for example, the column of label 1) can be calculated. By performing a multiple comparison test with the label used as the benchmark as other labels (for example, label 2), the size relationship between the label and multiple other labels (for example, label 1 and labels 3 to 6) based on the other label (for example, label 2) can be evaluated, and the p-value of each cell in a column of the p-value table (for example, the column of label 2) can be calculated.

[0058] According to the multiple comparison test, if the p-value in the p-value table for cells intersecting the reference label column and the rows of non-reference labels other than the reference label is 0.05 or less, the non-reference label score is concluded to be smaller than the reference label score with 95% confidence. Furthermore, if the p-value for all cells in the reference label column is 0.05 or less, the reference label score is considered the highest.

[0059] Using this multiple comparison test method, with the correct label set as the reference label, the knowledge transfer progress determination unit 12 determines, within the intermediate model output by the knowledge transfer learning unit 11, whether the reference likelihood of the learning data to which the reference label (i.e., the correct label) is assigned is greater than the likelihood of multiple learning data to which non-reference labels other than the reference label are assigned, and outputs the determination result. If the reference label is label 1, the non-reference labels correspond to labels 2 through 6.

[0060] (Knowledge Transfer Repeated Judgment Unit)

[0061] The knowledge transfer repetition determination unit 13 determines whether to iterate knowledge transfer learning based on the determination result output by the knowledge transfer progress determination unit 12. More specifically, if the baseline likelihood of the reference label is greater than the likelihoods of all non-reference labels other than the reference label, the knowledge transfer repetition determination unit 13 determines not to iterate knowledge transfer learning. If not (if not greater than), the knowledge transfer repetition determination unit 13 determines to iterate knowledge transfer learning. The knowledge transfer repetition determination unit 13 outputs the determination result of whether to iterate knowledge transfer learning.

[0062] (Knowledge transfer result output section)

[0063] When the knowledge transfer result output unit 14 receives the judgment result that the knowledge transfer learning is not repeated from the knowledge transfer repetition judgment unit 13, it regards the intermediate model as the learned model after the knowledge transfer is completed, and outputs the data of the learned model after the knowledge transfer as the knowledge transfer result to the storage device 4.

[0064] Next, refer to Figure 2A and Figure 2B , the hardware structure example of the transfer learning device 1 is described. The functional part of the transfer learning device 1 is realized by a processing circuit. That is, the transfer learning device 1 has a processing circuit for performing various operations such as knowledge transfer learning, knowledge transfer progress determination, knowledge transfer repetition determination, and knowledge transfer result output. The processing circuit can be Figure 2A The dedicated processing circuit 100a shown may also be Figure 2B A computer is shown having a processor 100b that executes a program stored in a memory 100c.

[0065] When the processing circuitry is a dedicated processing circuit 100a, the dedicated processing circuit 100a is, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (application specific integrated circuit), an FPGA (field-programmable gate array), or a combination thereof. The functions of the knowledge transfer learning unit 11, the knowledge transfer progress determination unit 12, the knowledge transfer iteration determination unit 13, and the knowledge transfer result output unit 14 can be implemented by multiple different processing circuits, or the functions of each unit can be implemented collectively by a single processing circuit.

[0066] When the processing circuitry is a processor 100b, the functions of the knowledge transfer learning unit 11, the knowledge transfer progress determination unit 12, the knowledge transfer repetition determination unit 13, and the knowledge transfer result output unit 14 are implemented by software, firmware, or a combination of software and firmware. Software and firmware are described as programs and stored in the memory 100c. The processor 100b implements the functions of each unit by reading and executing the program stored in the memory. In other words, the memory 100c has a program (transfer learning program) for storing steps such as the steps of performing knowledge transfer learning, the steps of determining the progress of knowledge transfer, the steps of determining the repetition of knowledge transfer, and the steps of outputting knowledge transfer results. Here, examples of the memory 100c include non-volatile or volatile semiconductor memories such as RAM (random access memory), ROM (read-only memory), flash memory, EPROM (erasable programmable read only memory), EEPROM (electrically erasable programmable read-only memory), magnetic disks, floppy disks, optical disks, compact disks, mini disks, and DVDs.

[0067] Furthermore, the functions of the knowledge transfer learning unit 11, the knowledge transfer progress determination unit 12, the knowledge transfer repetition determination unit 13, and the knowledge transfer result output unit 14 may be partially implemented by dedicated hardware and partially implemented by software or firmware. Thus, the processing circuit can implement the aforementioned functions through hardware, software, firmware, or a combination thereof.

[0068] <Action>

[0069] Next, refer to Figure 7 , the operation of the transfer learning device 1 will be described. In step S1, the knowledge transfer learning unit 11 performs knowledge transfer learning based on the learned model related to the first type of data using learning data related to a second type of data different from the first type, and generates an intermediate model based on the learned model.

[0070] In step S2, the knowledge transfer progress determination unit 12 uses a multiple comparison test to determine whether the baseline likelihood of the learning data with the baseline label is greater than the likelihoods of multiple learning data with non-baseline labels other than the baseline label, within the intermediate model generated by the knowledge transfer learning unit 11. The knowledge transfer progress determination unit 12 outputs the determination result.

[0071] In step S3, the knowledge transfer iteration determination unit 13 determines whether to iterate knowledge transfer learning based on the determination result output by the knowledge transfer progress determination unit 12. More specifically, if the baseline likelihood of the reference label is greater than the likelihoods of all non-reference labels other than the reference label, the knowledge transfer iteration determination unit 13 determines that knowledge transfer learning should not be iterated. Otherwise (if it is not greater than), the knowledge transfer iteration determination unit 13 determines that knowledge transfer learning should be iterated. If it is determined that it is iterated, the process returns to step S1. If it is determined that it is not iterated, the process proceeds to step S4.

[0072] In step S4, when the knowledge transfer result output unit 14 receives the determination result of non-repetition made by the knowledge transfer repetition determination unit 13, it regards the intermediate model as a learned model after knowledge transfer in which knowledge transfer is completed, and outputs the data of the learned model after knowledge transfer as the knowledge transfer result.

[0073] The transfer learning device 1 described above can evaluate, for each non-reference label, the degree of learning progress of learning data with multiple non-reference labels relative to learning data with reference labels. This allows for identification of learning data with delayed progress in transfer learning, providing a basis for focusing on learning with delayed progress in transfer learning. Consequently, the learning time for transfer learning can be shortened.

[0074] Implementation method 2.

[0075] <Structure>

[0076] Next, refer to Figures 8 to 9C Next, a transfer learning device 1' according to Embodiment 2 of the present disclosure will be described. This transfer learning device 1' according to Embodiment 2 differs from the transfer learning device 1 according to Embodiment 1 in that it includes a learning data selection unit 15. In response to the addition of learning data selection unit 15, the functionality of a knowledge transfer progress determination unit 12' included in the transfer learning device 1' differs slightly from that of the knowledge transfer progress determination unit 12 in the transfer learning device 1.

[0077] In the transfer learning device 1' of Embodiment 2, the knowledge transfer progress determination unit 12' determines the progress of knowledge transfer learning by determining whether there is learning data with delayed progress in knowledge transfer. This determination is performed by identifying non-reference labels with p-values less than a predetermined value based on the results of a multiple comparison test. Thus, if the knowledge transfer progress determination unit 12' does not determine that the reference likelihood is greater than the likelihoods of multiple learning data items to which non-reference labels are assigned, it identifies non-reference labels with p-values less than the predetermined value based on the results of the multiple comparison test.

[0078] The learning data selection unit 15 is a functional unit that, when a determination is made that the baseline likelihood is not greater than the likelihoods of multiple learning data items each labeled with a non-baseline label, selects learning data that contributes to the determination that the baseline likelihood is not greater than the likelihoods of multiple learning data items. More specifically, the learning data selection unit 15 selects (i.e., obtains) learning data labeled by the knowledge transfer progress determination unit 12' from the storage device 3 and provides the selected learning data to the knowledge transfer learning unit 11. The knowledge transfer learning unit 11 performs transfer learning using the learning data selected by the learning data selection unit 15.

[0079] For example, refer to Figure 9A To explain, the knowledge transfer progress determination unit 12' calculates the likelihood of the depth image in rounds 26 to 35. Figure 9A In the p-value table, label 6 is identified from column C1 with reference label number 1. Label 6 is identified by checking whether the p-value is a predetermined value, such as 0.05 or less. Learning data selection unit 15 selects learning data with label 6 from storage device 3 and provides the selected learning data to knowledge transfer learning unit 11.

[0080] exist Figure 9A In the example, if the p-values for labels 1 to 5 are below a predetermined value, it indicates that learning for labels 1 to 5 has progressed sufficiently. On the other hand, if the p-value for label 6 is greater than the predetermined value, it indicates that learning for label 6 has not progressed sufficiently. Therefore, label 6, whose learning progress is lagging behind the other labels, is prioritized for learning.

[0081] By focusing on the depth image of label 6, such as Figure 9B and Figure 9C As shown, the p-value of label 6 decreases sharply. Figure 9B This is the p-value table calculated based on the likelihood of the depth image in rounds 27 to 36. Figure 9C This is the p-value table calculated based on the likelihood of the depth image in rounds 28 to 37. By focusing on the depth image of label 6, the p-value of label 6 is Figure 9B Column C2 is reduced to 0.33, Figure 9C The value in column C3 decreases to 0.00.

[0082] By focusing on learning data with a delayed learning progress in this manner, the number of times of learning for data with a delayed learning progress can be reduced, thereby shortening the overall learning time required for transfer learning.

[0083] The hardware of the transfer learning device 1' is the same as the hardware configuration example of the transfer learning device 1 of the first embodiment, and can be Figure 2A or Figure 2BThis can be achieved by using the structure example of .

[0084] <Action>

[0085] Next, refer to Figure 10 , the operation of the transfer learning device 1 is explained. Steps S1, S3 and S4 are Figure 7 The steps are the same as in , so repeated descriptions are omitted.

[0086] In step S2', the knowledge transfer progress determination unit 12 determines whether there is learning data with delayed progress of knowledge transfer. If there is delayed learning data, the process proceeds to step S5, and if there is no delayed learning data, the process proceeds to step S3.

[0087] In step S5, the learning data selection unit 15 selects the learning data determined by the knowledge transfer progress determination unit 12 to be delayed in progress of knowledge transfer from the storage device 3, and supplies the selected learning data to the knowledge transfer learning unit 11. The process then returns to step S1.

[0088] Furthermore, the embodiments may be combined, or each embodiment may be modified or omitted as appropriate.

[0089] Industrial applicability

[0090] The transfer learning device disclosed herein can be used as a device for transferring and utilizing knowledge related to model assets that have been fully learned through machine learning, and can be used, for example, in technical fields such as image analysis.

[0091] Label Description

[0092] 1(1'): transfer learning device; 2: storage device; 3: storage device; 4: storage device; 11: knowledge transfer learning unit; 12(12'): knowledge transfer progress determination unit; 13: knowledge transfer repetition determination unit; 14: knowledge transfer result output unit; 15: learning data selection unit; 100a: processing circuit; 100b: processor; 100c: memory.

Claims

1. A transfer learning device, wherein: The transfer learning device has: a knowledge transfer learning unit that performs knowledge transfer learning based on a learned model related to data of a first category using learning data related to a second category different from the first category, and generates an intermediate model based on the learned model; as well as The knowledge transfer progress determination unit determines, in the generated intermediate model, whether the reference likelihood of the learning data to which the reference label is added is greater than the likelihoods of a plurality of learning data to which non-reference labels other than the reference label are added, by a multiple comparison test.

2. The transfer learning device according to claim 1, wherein: The transfer learning device further includes a learning data selection unit that selects learning data that contributes to not determining that the reference likelihood is greater than the likelihoods of the plurality of learning data, when the reference likelihood is not determined to be greater than the likelihoods of the plurality of learning data.

3. The transfer learning device according to claim 1 or 2, wherein: The transfer learning device further includes a knowledge transfer repetition determination unit configured to determine to terminate the knowledge transfer learning when determining that the reference likelihood is greater than the likelihoods of the plurality of learning data.

4. A transfer learning method, performed by a transfer learning device having a knowledge transfer learning unit and a knowledge transfer progress determination unit, wherein: The transfer learning method includes: The knowledge transfer learning unit performs knowledge transfer learning based on a learned model related to data of a first category using learning data related to a second category different from the first category, and generates an intermediate model based on the learned model; as well as The knowledge transfer progress determination unit determines the step of the knowledge transfer progress. In the generated intermediate model, a multiple comparison test is used to determine whether the baseline likelihood of the learning data with the baseline label attached is greater than the likelihood of multiple learning data with non-baseline labels other than the baseline label attached.

5. A transfer learning program wherein: The transfer learning program enables the computer to: a function of performing knowledge transfer learning, based on a learned model related to data of a first category, using learning data related to data of a second category different from the first category, and generating an intermediate model based on the learned model; as well as The function of determining the progress of knowledge transfer determines, in the generated intermediate model, whether the baseline likelihood of the learning data with the baseline label attached is greater than the likelihood of multiple learning data with non-baseline labels other than the baseline label attached by multiple comparison tests.

Images