JP7361999B1 - Machine learning device, machine learning system, machine learning method, and machine learning program - Google Patents

Abstract

The machine learning device (10) performs normal learning in which a learning model for inferring a region of interest (B) of input data is generated based on learning data (L), and transfer learning in which the learning model is updated. a learning section (11, 15a); a pre-update region of interest storage section (13) that generates a region of interest of input data using a learning model during the normal learning, and uses the generated region of interest as a pre-update region of interest; During or before transfer learning, the area of interest of the input data is updated from the area of interest before update (Bi) to the area of interest after update (Ai) using the input person's knowledge. A region-of-interest update section (18) performs update processing to store the updated region-of-interest storage section (14), and a region-of-interest update evaluation section (15) calculates a similarity distance (di) indicating a change in the region of interest due to the update process. ), and a data dynamic selection unit (16) that dynamically selects input data to be output to the region of interest update unit (18) based on the similarity distance during or before transfer learning.

Description

The present disclosure relates to a machine learning device, a machine learning system, a machine learning method, and a machine learning program.

In deep learning, improvement of prediction accuracy in inference using a learning model is achieved by introducing an Attention Map that indicates a region of interest that a learning model should pay attention to. The region of interest is acquired by learning the learning model, but by performing transfer learning (that is, human-in-the-loop learning) that manually corrects the region of interest, the performance of the learning model can be improved (predicted). (for example, see Patent Documents 1 and 2).

JP2021-022368A International Publication No. 2020/085336

However, in transfer learning, it is not clear which region of data out of a large amount of data should be manually updated to efficiently improve the performance of a learning model. Therefore, in transfer learning that manually updates a region of interest in a large amount of data such as a video, it is desirable to improve the performance of a learning model with less effort (that is, to improve the efficiency of transfer learning).

The present disclosure aims to provide a machine learning device, a machine learning system, a machine learning method, and a machine learning program that make it possible to efficiently improve the performance of a learning model generated in transfer learning.

The machine learning device of the present disclosure performs normal learning in which a learning model for inferring a region of interest of input data is generated based on learning data collected in advance, and transfer learning in which the learning model is updated. and a generation process of generating the region of interest of the input data using the learning model during the normal learning, and storing the generated region of interest in a pre-update region of interest storage unit as a pre-update region of interest. During or before the transfer learning, update the region of interest of the input data from the pre-update region of interest to the post-update region of interest using the inputted person's knowledge and store it in the post-update region of interest storage unit. a region of interest update unit that performs an update process to save; a region of interest update evaluation unit that calculates a similarity distance indicating a change in the region of interest due to the update process; and a data dynamic selection section that dynamically selects the input data to be output to the region of interest update section based on the above.

The machine learning method of the present disclosure is a machine that performs normal learning in which a learning model for inferring a region of interest in input data is generated based on learning data collected in advance, and transfer learning in which the learning model is updated. A method implemented by a learning device, wherein during the normal learning, the region of interest of the input data is generated using the learning model, and the generated region of interest is used as the region of interest before update as the region of interest before update. A generation process is performed to store the data in a region storage unit, and during or before the transfer learning, the region of interest of the input data is updated from the pre-update region of interest to the post-update region of interest using the inputted person's knowledge. performing an update process to store the area of interest in the updated area of interest storage unit; calculating a similarity distance indicating a change in the area of interest due to the update process; and calculating the similarity distance during or before the transfer learning. Dynamically selecting the input data to be subjected to the transfer learning based on the method.

By using the machine learning device, machine learning system, machine learning method, and machine learning program of the present disclosure, it is possible to efficiently improve the performance of a learning model generated in transfer learning.

1 is a functional block diagram schematically showing the configuration of a machine learning device according to Embodiment 1 (including the configuration of a task learning unit that is a supervised learning unit); FIG. 1 is a functional block diagram schematically showing the configuration of a machine learning device according to Embodiment 1 (including the configuration of a dynamic selection/update section). FIG. FIG. 3 is a diagram illustrating the operation of a region of interest update evaluation unit of the machine learning device. FIG. 3 is a diagram illustrating the operation of a region of interest update evaluation unit of the machine learning device. It is a figure showing operation of a data dynamic selection part of a machine learning device. (A) and (B) are diagrams showing the operation of a search utilization adjustment unit of the machine learning device. (A) and (B) are diagrams showing the operation of a search utilization adjustment unit of the machine learning device. 3 is a flowchart showing the operation of the machine learning device according to the first embodiment. 1 is a diagram illustrating an example of a hardware configuration of a machine learning device according to Embodiment 1. FIG. FIG. 2 is a functional block diagram schematically showing the configuration of a machine learning device according to a second embodiment. 7 is a flowchart showing the operation of the machine learning device according to Embodiment 2. FIG. 3 is a functional block diagram schematically showing the configuration of a machine learning device according to Embodiment 3. FIG. 7 is a flowchart showing the operation of the machine learning device according to Embodiment 3. FIG. 7 is a functional block diagram schematically showing the configuration of a machine learning device and a machine learning system according to a fourth embodiment. 12 is a flowchart showing operations of a machine learning device and a machine learning system according to Embodiment 4.

Below, a machine learning device, a machine learning system, a machine learning method, and a machine learning program according to embodiments will be described with reference to the drawings. The following embodiments are merely examples, and the embodiments can be combined as appropriate and each embodiment can be changed as appropriate.

Embodiment 1.
FIG. 1 is a functional block diagram schematically showing the configuration of a machine learning device 10 according to the first embodiment (including the configuration of a task learning unit 11 that is a supervised learning unit). The machine learning device 10 is, for example, a learning device that can perform transfer learning that incorporates human knowledge. The machine learning device 10 is, for example, a computer. The machine learning device 10 is a device that can implement the machine learning method according to the first embodiment. The machine learning device 10 includes a task learning section 11 and a dynamic selection/updating section 19. The machine learning device 10 is connected to a learning data storage section 12 , a pre-update region of interest storage section 13 , and a post-update region of interest storage section 14 . The machine learning device 10, the learning data storage section 12, the pre-update region of interest storage section 13, and the post-update region of interest storage section 14 constitute a machine learning system. The learning data storage unit 12, the pre-update area of interest storage unit 13, and the post-update area of interest storage unit 14 may be provided in a storage device inside the machine learning device 10. The learning data storage unit 12, the pre-update area of interest storage unit 13, and the post-update area of interest storage unit 14 may be different storage areas of a common storage device.

The task learning section 11 includes a learning model generation section 11a and a learning model storage section 11b. The learning model generation unit 11a acquires learning data L including input data and correct answer data, and uses the learning data L to generate a focus from actual input data (for example, thermal image data input in the inference process). A learning model M for inferring a region is generated and stored in the learning model storage unit 11b. When performing transfer learning, the learning model generation unit 11a generates a learning model M using the learning data L and the updated region of interest A, and stores it in the learning model storage unit 11b. Note that the machine learning device 10 may include an inference unit that acquires input data and outputs an inference result obtained from the input data using the learning model M. The machine learning device 10 in this case is a machine learning/inference device. Further, the inference unit may be provided in a computer different from the machine learning device 10.

The learning data storage unit 12 stores learning data L collected in advance according to the purpose of machine learning, and provides the learning data L to the task learning unit 11. The learning data storage unit 12 may be provided in a storage device external to the machine learning device 10.

The dynamic selection/update unit 19 acquires the learning model M from the task learning unit 11 , acquires the pre-update region of interest B from the pre-update region of interest storage unit 13 , and performs task learning via the post-update region of interest storage unit 14 . The updated region of interest is given to the section 11.

When performing transfer learning, the dynamic selection/updating unit 19 generates an updated region of interest obtained by manually (i.e., user 50 as an intervener) modifying the region of interest generated by the learning model M. Ai is given to the updated target area storage unit 14. The updated region of interest A stored in the updated region of interest storage section 14 is given to the task learning section 11 . Here, “Ai” indicates the updated area of interest of the i-th data, that is, the selected updated area of interest, and “A” indicates all the updated areas of interest stored in the updated area of interest storage unit 14. Indicates the subsequent region of interest. Note that i is information for identifying data, and is, for example, a positive integer.

FIG. 2 is a functional block diagram schematically showing the configuration of the machine learning device 10 (including the configuration of the dynamic selection/update section 19). The dynamic selection/update section 19 includes a region of interest update evaluation section 15 , a data dynamic selection section 16 , a search utilization adjustment section 17 , and a region of interest update section 18 .

The machine learning device 10 performs normal learning in which a learning model M for inferring a region of interest in input data is updated based on learning data L collected in advance, and transfer learning in which the learning model M is updated. (similarity learning unit 15a and task learning unit 11). During normal learning, the dynamic selection/update unit 19 generates a region of interest of input data using the learning model generated by the learning section, and stores the generated region of interest as a region of interest before update as a region of interest before update. Performs generation processing to be stored in the section 13, and updates the region of interest of the input data from the pre-update region of interest Bi to the post-update region of interest Ai using the inputted person's knowledge during or before transfer learning. It has a region-of-interest update section 18 that performs an update process to store the region of interest in the region-of-interest storage section 14 . In addition, the dynamic selection/update unit 19 is connected to a region of interest update evaluation unit 15 that calculates a similarity distance d ⁱ indicating a change in the region of interest due to the update process, and a region of interest update evaluation unit 15 that calculates a similarity distance d ⁱ during or before transfer learning. The data dynamic selection section 16 dynamically selects the input data to be output to the region of interest updating section 18 based on the following.

The learning data storage unit 12 stores learning data L collected according to a specific task. For example, in the case of an image classification task, the learning data storage unit 12 stores a set of image data (for example, a human image, a medical image, etc.) and a class number corresponding to each image (i.e., a learning data set). . The learning data L is not limited to data related to images, but may be data related to natural language, table data, or the like.

The task learning unit 11 receives learning data L as input, and learns a learning model M based on the learning data L. During transfer learning, the task learning unit 11 acquires the learning data L and the updated region of interest A as input, and outputs the pre-updated region of interest B during or after learning. That is, the task learning unit 11 outputs the pre-update region of interest B either during or after learning the learning model M. Furthermore, in order for the task learning unit 11 to output the pre-update region of interest B during learning, a known method called Attention Branch Network (ABN) may be employed. Furthermore, in order to output the pre-update region of interest B after learning, the task learning unit 11 generates the region of interest after the fact using Local Interpretable Model-agnostic Explanation (LIME) or Class Activation Mapping (CAM). Any known method may be used.

The pre-update region of interest storage unit 13 stores a region of interest (pre-update region of interest B) corresponding to each data. The updated area of interest storage unit 14 stores an updated area of interest A corresponding to each piece of data. Note that the selected pre-update region of interest is indicated by Bi, and the selected post-update region of interest is indicated by Ai. Although the data structure of the pre-update region of interest B and the post-update region of interest A is free, it is desirable that the data structure be interpretable and editable by humans (that is, the user 50). An example of the data structure of the pre-update region of interest B and the post-update region of interest A is a heat map or the like. The pre-update area of interest B and the post-update area of interest A may be stored in the same storage device.

The region of interest update unit 18 includes or is connected to a user interface (UI) for a user 50 who is an interventionist to perform update work. The UI is, for example, an input device such as a keyboard, mouse, touch panel, or paint tool for drawing. For example, in the image displayed on the display, the region of interest update unit 18 deletes a part of the region of interest (i.e., shrinks it), adds the region of interest (i.e., expands it), or both deletes and adds the region of interest. (including, for example, moving the region of interest). At this time, it is possible to carry out a process of weighting each of the plurality of regions of interest. Furthermore, a configuration may be provided to introduce a threshold value for making a determination regarding weighted data and to perform interpolation processing using weighted data. A process of weighting each of a plurality of regions of interest is shown in, for example, Patent Document 2. After the update of the region of interest is completed, the updated region of interest is sent to the updated region of interest storage section 14 and the region of interest update evaluation section 15 as the updated region of interest Ai.

The region of interest update evaluation unit 15 calculates the similarity distance d ⁱ between the region of interest before update Bi and the region of interest after update Ai, and evaluates the validity of the region of interest before update Bi. The simplest method for calculating the similarity distance d ⁱ for the i-th data is to calculate the squared error for the raw data of the pre-update region of interest and the post-update region of interest. However, in the case of unstructured data such as images and natural language, it is not possible to capture semantic differences, so intermediate features acquired by the learning model M are used instead of raw data to calculate the similarity distance d ⁱ Calculate. Any distance function such as cosine similarity or Mahalanobis distance can be used to calculate the similarity distance d ⁱ . Note that a specific example of the operation of the region of interest update evaluation unit 15 will be described later using FIGS. 3 and 4.

The data dynamic selection unit 16 performs the next update based on the update history of the user 50, using the correspondence between the intermediate feature of the target area before update and the similarity distance d ⁱ obtained by the target area update evaluation unit 15. Dynamically select the desired data. First, the similarity learning unit 15a newly learns a learning model using the above correspondence as input and output. At this time, the similarity learning unit 15a learns a Bayesian model such as a Gaussian process regression model so that not only the average Av but also the deviation D is output for predicting the similarity distance d ⁱ . Finally, the data dynamic selection unit 16 selects data points that maximize the acquisition function based on the average Av and deviation D of the similarity distance d ⁱ . This process can be used in conjunction with a static data selection method, and the data selected at one time may be a single piece of data or multiple pieces of data. Note that a specific example of the operation of the data dynamic selection section 16 will be described later using FIG. 5.

By adjusting the hyperparameter H of the acquisition function, the search utilization adjustment unit 17 considers the balance between "low validity of the region of interest before update" and "low similarity with past updated data". Select the data. A specific example of the operation of the search utilization adjustment unit 17 will be described later using FIGS. 6 and 7. The search/utilization adjustment unit 17 provides the data dynamic selection unit 16 with a hyperparameter H indicating whether emphasis is placed on search or utilization of the learning data L acquired in advance. The data dynamic selection unit 16 dynamically selects input data to be output to the region of interest updating unit 18 based on the similarity distance d ⁱ and the hyperparameter H during transfer learning. During transfer learning, the search and utilization adjustment unit 17 preferably first sets the hyperparameter to a value that emphasizes search, and then gradually changes the hyperparameter from a value that emphasizes search to a value that emphasizes utilization.

FIG. 3 is a diagram showing the operation of the region of interest update evaluation unit 15. In transfer learning, the area of interest after update Ai (areas 62a, 62b or area 63 in FIG. 3) obtained by updating the area of interest Bi before update (areas 61a, 61b in FIG. 3) by the area of interest updating unit 18 and the learning The difference between the region of interest (regions 64a and 64b in FIG. 3) newly generated by model M becomes a penalty. Therefore, in order to improve the performance of the learning model M (i.e., efficiently improve prediction accuracy and interpretability) by performing transfer learning, the validity of the pre-update region of interest Bi is low (i.e., It is desirable to update data that undergoes large changes due to update work (that is, preferentially). In the first embodiment, by dynamically selecting data with low validity in the pre-update region of interest Bi, it is possible to efficiently improve the performance of the learning model M by updating a smaller number of data.

In FIG. 3, when the updated region of interest Ai (correct data) is Example 1, as can be seen by comparing regions 62a and 62b with regions 64a and 64b, which are regions of interest generated by the learning model M, the region 62a , 62b and the regions 64a, 64b are located at similar positions and have similar sizes. In this case, the similarity distance d ⁱ between the two is small, the penalty that is the difference between the two is small, and the validity of the pre-update region of interest Bi is high, so it is necessary to manually update the pre-update region of interest Bi. gender is low.

In FIG. 3, when the updated region of interest Ai (correct data) is Example 2, as can be seen by comparing the region 63 and regions 64a and 64b, which are the regions of interest generated by the learning model M, the region 63 and the region 64b are similar to each other, but there is no region corresponding to the region 64a in the updated region of interest Ai of Example 2. In this case, the similarity distance d ⁱ between the two is large, the penalty that is the difference between the two is large, and the validity of the pre-update region of interest Bi is low, so it is necessary to manually update the pre-update region of interest Bi. The quality is high.

As can be seen from FIG. 3, in the first embodiment, during transfer learning, the learning model M is efficiently performance can be improved.

FIG. 4 is a diagram showing the operation of the region-of-interest update evaluation unit 15 of the machine learning device 10. FIG. 4 shows that the pre-update region of interest Bi (= x ⁱ _before ), which is local data including “horses and people”, is changed by the region of interest updating unit 18 into the post-update region of interest, which is local data including “horses”. Ai (= x ⁱ _after ), and by the learning model M, intermediate features h ⁱ _before of the area of interest before update Bi (= x ⁱ _before ) and intermediate features of the area of interest after update Ai (= x ⁱ _after ) h ⁱ _after can be obtained, and the similarity distance d ⁱ can be obtained from these intermediate features. In other words, the region of interest update evaluation section 15 quantifies the change in the region of interest due to the update work (i.e., update processing) performed by the region of interest update section 18 using the similarity distance d ⁱ between the intermediate features extracted by the learning model M. . The similarity distance d ⁱ is a value representing the validity of the pre-update region of interest Bi (=x ⁱ _before ). The intermediate feature is a feature of the data (local data) of the region of interest extracted from the image data using a mask. The similarity distance d ⁱ is the intermediate feature h ⁱ _before of the region of interest before update Bi (= x ⁱ _before ) and the intermediate feature h ⁱ of the region of interest after update Ai (= x ⁱ _after ) given from the region of interest updating unit 18. The distance function f based on _after can be expressed by the following equation (1). Here, the distance function f is, for example, cosine similarity.

FIG. 5 is a diagram showing the operation of the data dynamic selection section 16. The data dynamic selection unit 16 estimates the validity of the region of interest before updating (that is, the expected value and variance of the similarity distance d ⁱ ) for the unupdated data, and performs Bayesian optimization on the data points to be updated next. Select by the following method. For example, the data dynamic selection unit 16 calculates an acquisition function for all unupdated data and selects the data with the largest value.

FIGS. 6A and 6B are diagrams showing the operation of the search utilization adjustment unit 17. FIG. 6(A) shows an example of Bayesian optimization. FIG. 6(B) shows an example of UCB (Upper Confidence Bound) which is an acquisition function. By adjusting the UCB hyperparameters, the search utilization adjustment unit 17 considers the balance between "low validity of the region of interest before update" and "low similarity with past updated data". Select data.

In Bayesian optimization, the search utilization adjustment unit 17 uses unupdated data that maximizes the acquisition function as data to be presented to the user next. At this time, by changing the hyperparameter β _t of the acquisition function α _t (x), it is possible to control the balance between exploration (when emphasizing data with a large deviation) and utilization (when emphasizing data with a large average). can. The user 50 adjusts the hyperparameter β _t using a UI such as a slider, and proceeds with the update work while balancing exploration and utilization. For example, by gradually decreasing the value of the hyperparameter β _t , the importance of exploration can be increased at the beginning of the task (that is, the importance of utilization can be decreased), and the importance of utilization can be increased at the end of the task ( In other words, data can be selected to reduce the importance of the search. The operation of the search utilization adjustment unit 17 can be expressed by the following equation (2).

In equation (2), x represents a Gaussian random variable, α _t (x) represents Gaussian Process UCB (GP-UCB) as an acquisition function, and x _next represents α _t (x) indicates the point where is maximum. Further, μ _t (x) indicates the predicted average of x, β _t indicates the hyperparameter, and σ _t (x) indicates the predicted deviation of x. If β _t approaches 0, the setting emphasizes search, and if β _t becomes large, the setting emphasizes utilization. Note that the subscript t indicates the number of iterations, and the subscript t is omitted in the figure.

FIGS. 7A and 7B are diagrams showing the operation of the search utilization adjustment unit 17. Adjustment of hyperparameters cannot be done intuitively, so it is not easy for humans to do this adjustment. If you place extreme emphasis on utilization (in Figure 7 (A), if the hyperparameter β _t approaches 0), you can use unupdated data that has a high degree of similarity to updated data (for example, in the case of a video, before and after an updated frame) frame) will be selected, making the task of updating the region of interest redundant. On the other hand, if the search is extremely important (in FIG. 7A, if the hyperparameter β _t is made excessively large), data with high validity (that is, data that the user 50 does not need to update) will be The possibility of being selected increases, and the performance of the generated learning model cannot be efficiently improved. Therefore, the search utilization adjustment unit 17 introduces a new hyperparameter β _t in order to exclude these inappropriate unupdated data from being selected. Specifically, as shown in FIG. 7(B), for each of the predicted average μ _t (x) and the predicted deviation σ _t (x) of the similarity distance d ⁱ , unupdated items to be excluded from selection are Determine the percentage of data. In FIG. 7B, the upper limit of the hyperparameter β _t is determined by excluding data in the lower a% (a is a set value) of the predicted average μ _t (x). Further, in FIG. 7B, the lower limit of the hyperparameter β _t is determined by excluding data in the top b% (b is a set value) of the prediction deviation σ _t (x). This is equivalent to limiting the range of possible values of the hyperparameter β _t . This determines the range of data selected by the acquisition function α _t (x).

FIG. 8 is a flowchart showing the operation of the machine learning device 10. First, the learning unit performs supervised learning using the learning data L to generate a learning model M (step S101). The dynamic selection/updating unit 19 generates a region of interest for the learning data L using the learning model M (step S102). Initially, since this is the first update (YES in step S103), the dynamic selection/update unit 19 statically selects the data (step S108), updates the region of interest (step S106), and updates the update result. Evaluation is performed (step S107).

Next, if the update is to be continued (YES in step S109), the dynamic selection/update unit 19 adjusts the hyperparameters (step S104) and dynamically selects the data (step S105). The region of interest of the acquired data is updated (step S106) and the update results are evaluated (step S107).

When the dynamic selection/updating unit 19 completes updating the dynamically selected data, it does not continue updating (NO in step S109) and determines whether to perform transfer learning (step S110). When performing transfer learning (YES in step S110), the task learning unit 11 performs supervised learning using the learning data L and the updated region of interest A to update the learning model M (step S101). After that, the dynamic selection/updating unit 19 generates a region of interest for all data using the updated learning model M (step S102). Next, the dynamic selection/update unit 19 adjusts hyperparameters (step S104), dynamically selects data (step S105), updates the region of interest of the selected data (step S106), and evaluates the update results. (Step S107) is repeated until the update continuation is ended (NO in step S109). Further, the dynamic selection/update unit 19 repeatedly performs the processes of steps S101 to S107 and S109 until the transfer learning is finished (NO in step S110).

FIG. 9 is a diagram showing an example of the hardware configuration of the machine learning device 10. The machine learning device 10 includes a processor 101, a memory 102 that is a volatile storage device, a nonvolatile storage device 103 such as a hard disk drive (HDD) or a solid state drive (SSD), and an interface 104. . The memory 102 is, for example, a semiconductor memory such as a RAM (Random Access Memory). The machine learning device 10 may include a communication device that communicates with an external device.

Each function of the machine learning device 10 is realized by a processing circuit. The processing circuit may be dedicated hardware or may be the processor 101 that executes a program (for example, a machine learning program according to the embodiment) stored in the memory 102. The processor 101 may be any one of a processing device, an arithmetic device, a microprocessor, a microcomputer, and a DSP (Digital Signal Processor).

When the processing circuit is dedicated hardware, the processing circuit is, for example, an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

When the processing circuit is the processor 101, the machine learning method is performed by software, firmware, or a combination of software and firmware. Software and firmware are written as programs and stored in memory 102. Processor 101 can implement the machine learning method according to the first embodiment by reading and executing the program stored in memory 102.

Note that the machine learning device 10 may have a part implemented by dedicated hardware and another part by software or firmware. In this way, the processing circuit can implement each of the above-mentioned functions using hardware, software, firmware, or any combination thereof.

Interface 104 is used to communicate with other devices. The interface 104 is connected to an external storage device 105, a display 106, an input device 107 as a user operation unit, and the like.

As explained above, if the machine learning device 10 according to the first embodiment is used, in transfer learning, the difference between the pre-update region of interest Bi and the post-update region of interest Ai newly generated by the learning model M can be It will be a penalty. Therefore, in order to promote accuracy improvement through transfer learning, it is desirable to update data with low validity in the pre-update region of interest B in a focused manner. According to the machine learning device 10 according to the first embodiment, data with low validity (that is, data that changes significantly due to update work and needs to be updated) in the pre-update focused area B is dynamically Because of this selection, the performance of the learning model can be efficiently improved by updating a smaller amount of data.

Embodiment 2.
FIG. 10 is a functional block diagram schematically showing the configuration of the machine learning device 20 according to the second embodiment. In FIG. 10, components that are the same as or correspond to those shown in FIG. 2 are given the same reference numerals as those shown in FIG. The machine learning device 20 according to the second embodiment is different from the machine according to the first embodiment shown in FIG. This is different from the learning device 10. The task learning unit 21 and the unsupervised learning unit 22 learn a learning model M used for feature extraction of the pre-update region of interest in the region of interest update evaluation unit 15 . At this time, any learning model M having an encoder such as Principle Component Analysis (PCA) or Auto-Encoder (AE) may be used. Further, the unsupervised learning unit 22 may perform self-supervised learning.

Since the learning model M created by the task learning unit 21 selectively extracts features that are highly relevant to the task from the data, the learning model M ignores some of the features included in the updated region of interest A. there's a possibility that. When this learning model M is used in the region of interest update evaluation section 15, a distance that does not follow the manifold (that is, a distance that is "out-of-manifold") can be obtained. In order to avoid this problem, the learning model M created by the unsupervised learning section 22 is used as the feature extractor of the region of interest update evaluation section 15.

FIG. 11 is a flowchart showing the operation of the machine learning device 20. In FIG. 11, steps having the same contents as those shown in FIG. 8 are given the same reference numerals as those shown in FIG. The operation of the machine learning device 20 according to the second embodiment differs from the operation of the machine learning device 10 according to the first embodiment in that unsupervised learning is performed (step S201). In other respects, the operation of FIG. 11 is similar to that of FIG.

As described above, by using the machine learning device 20 according to the second embodiment, a general-purpose feature representation that is not task-specific can be obtained by unsupervised learning. Since this feature representation also includes features related to the updated region of interest, when used in the region of interest update evaluation section, it is possible to calculate an appropriate similarity distance.

Note that the second embodiment is the same as the first embodiment except for the above.

Embodiment 3.
FIG. 12 is a functional block diagram schematically showing the configuration of the machine learning device 30 according to the third embodiment. In FIG. 12, components that are the same as or correspond to those shown in FIG. 2 are given the same reference numerals as those shown in FIG. The machine learning device 30 according to the third embodiment differs from the embodiment shown in FIG. 2 in that it includes a data number reduction unit 31 that reduces input data supplied to the data dynamic selection unit 16 during transfer learning. This is different from the machine learning device 10 according to the first embodiment.

The data number reduction unit 31 reduces the number of data using unsupervised learning such as clustering. The data number reduction unit 31 selects a certain percentage of data from each class, or selects data in units of clusters. Hyperparameters used at this time, such as the number of clusters or data reduction ratio, may be set in advance, or may be changed based on a specific algorithm each time data is reduced.

In the data dynamic selection unit 16, creating a Bayesian model using all the data (each frame in the case of a video) and calculating an acquisition function for all the data requires a very large calculation cost. The amount of calculation in the case of Gaussian process regression is O(N ³ ) for the number of data N when expressed in "O notation". Therefore, in the machine learning device 30 according to the third embodiment, a data number reduction unit 31 is introduced that reduces the number of data using an algorithm such as random sampling, clustering, Kernel Interpolation for Scalable Structured Gaussian Processes (KISS-GP), etc. . Specifically, similar to the region of interest update evaluation section 15, the data number reduction section 31 uses the learning model M generated by the task learning section 11 to extract features of the region of interest B before update, and extracts intermediate features thereof. Execute unsupervised learning with input.

FIG. 13 is a flowchart showing the operation of the machine learning device 30. In FIG. 13, steps having the same contents as those shown in FIG. 8 are given the same reference numerals as those shown in FIG. The operation of the machine learning device 30 according to the third embodiment differs from the operation of the machine learning device 10 according to the first embodiment in that the process of reducing the number of data is performed (step S301). In other respects, the operation of FIG. 13 is the same as that of FIG.

As described above, by using the machine learning device 30 according to the third embodiment, the calculation cost in the data dynamic selection section 16 can be reduced, and the cost of machine power and the waiting time of the intervention person can be reduced.

Note that the third embodiment is the same as the first embodiment except for the above.

Embodiment 4.
FIG. 14 is a functional block diagram schematically showing the configuration of the machine learning device 40 and the configuration of the machine learning system according to the fourth embodiment. In FIG. 14, components that are the same as or correspond to those shown in FIG. 2 are given the same reference numerals as those shown in FIG. The machine learning system according to the fourth embodiment includes an effect measurement unit 41 that measures the degree of improvement in the inference process executed using the learning model M generated by the machine learning device 40, and a This machine learning system differs from the machine learning system according to the first embodiment shown in FIG. 2 in that it includes an update method determining unit 42 that determines hyperparameters related to updating a region.

The effect measurement unit 41 obtains, for example, prediction accuracy, interpretability, time required for each process, and cost required for resources in the inference process executed using the learning model M generated by the machine learning device 40, The degree of improvement of the obtained learning model M with respect to prediction accuracy, interpretability, time required for each process, and cost required for resources is measured. What is measured depends on the task and work environment (crowdsourcing, etc.). The degree of improvement of the learning model M does not need to be based on all of the prediction accuracy, interpretability, time required for each process, and cost required for resources; it may be one or a combination of two or more of these. good.

The update method determining unit 42 determines a hyperparameter H related to updating the region of interest based on the degree of improvement of the learning model M measured by the effect measuring unit 41, that is, the measured effect. The hyperparameter H may be determined on a rule-based basis or may be optimized. The determined hyperparameter H is automatically provided to the region of interest updating unit 18 and the search utilization adjustment unit 17 of the machine learning device 40. The determined hyperparameter H may be presented to the user 50 via a display to prompt the user 50 to change the hyperparameter H.

In updating the area of interest, multiple factors such as control parameters for exploration and utilization, UI design for updating the area of interest (e.g. granularity of the area of interest, etc.), number of data to be manually updated, number of interventionists, etc. are required. Hyperparameters exist. It is desirable that the hyperparameters be changed to appropriate values each time relearning is repeated. In the machine learning system according to the fourth embodiment, the degree of improvement of the learning model M in the inference process executed using the learning model M is measured, and the hyperparameter H is adjusted based on these values.

FIG. 15 is a flowchart showing the operation of the machine learning system according to the fourth embodiment. In FIG. 15, steps having the same contents as those shown in FIG. 8 are given the same reference numerals as those shown in FIG. In the operation of the machine learning system according to the fourth embodiment, the effect measuring unit 41 measures the degree of improvement (that is, the effect) of the learning model M (step S401), and the update method determining unit 42 measures the degree of improvement (that is, the effect) of the learning model M based on the measured effect. , differs from the operation in FIG. 8 in that the method of updating the region of interest (that is, the related hyperparameter H) is determined (step S402). In other respects, the operation of FIG. 15 is the same as that of FIG.

As explained above, there is a trade-off relationship between the load required for manual update work and its effect, but if the machine learning device 40 and machine learning system according to the fourth embodiment are used, a good balance between the two can be achieved. This makes it possible to perform updated updates.

Note that the fourth embodiment is the same as the first embodiment except for the above.

10, 20, 30, 40 machine learning device, 11 task learning unit (supervised learning unit), 11a learning model generation unit, 11b learning model storage unit, 12 learning data storage unit, 13 pre-update focused area storage unit, 14 update Post-focus area storage unit, 15 Focus area update evaluation unit, 15a Similarity learning unit (supervised learning unit), 16 Data dynamic selection unit, 17 Search utilization adjustment unit, 18 Focus area update unit, 19, 29, 39, 49 dynamic selection/update unit, 21 task learning unit (supervised learning unit), 22 unsupervised learning unit, 31 data number reduction unit, 41 effect measurement unit, 42 update method determination unit, 50 user, A, Ai, x ⁱ _after updated area of interest, B, Bi, x ⁱ _before updated area of interest, H, βt hyperparameter, d ⁱ similarity distance, L learning data, M learning model.

Claims (9)

a learning unit that performs normal learning that generates a learning model for inferring a region of interest of input data based on pre-collected learning data; and transfer learning that updates the learning model;
during the normal learning, generating the region of interest of the input data using the learning model, and performing a generation process of storing the generated region of interest in a pre-update region of interest storage unit as a pre-update region of interest; During or before the transfer learning, updating the region of interest of the input data from the pre-update region of interest to the post-update region of interest using the inputted person's knowledge, and storing the updated region of interest in the post-update region of interest storage unit. A region of interest update unit that performs processing;
a region of interest update evaluation unit that calculates a similarity distance indicating a change in the region of interest due to the update process;
a data dynamic selection unit that dynamically selects the input data to be output to the region of interest update unit based on the similarity distance during or before the transfer learning;
A machine learning device characterized by having. The machine learning device according to claim 1, wherein the learning unit is a supervised learning unit. The learning department is
an unsupervised learning section that performs the normal learning;
a supervised learning unit that performs the transfer learning;
The machine learning device according to claim 1, comprising: 4. The method according to claim 1, further comprising a data number reduction unit that reduces the input data supplied to the data dynamic selection unit during or before the transfer learning. Machine learning device. further comprising a search and utilization adjustment unit that provides the data dynamic selection unit with a hyperparameter indicating whether to emphasize search or utilization in the learning data acquired in advance;
The data dynamic selection unit dynamically selects the input data to be output to the region of interest updating unit based on the similarity distance and the hyperparameter during or before the transfer learning. The machine learning device according to any one of claims 1 to 4. The search utilization adjustment unit sets the hyperparameter to a value that emphasizes the search during or before the transfer learning, and gradually changes the hyperparameter from the value that emphasizes the search to the value that emphasizes the utilization. The machine learning device according to claim 5. A machine learning device according to claim 5 or 6,
an effect measurement unit that measures the degree of improvement in an inference process executed using the learning model generated by the machine learning device;
an update method determining unit that determines the hyperparameter related to updating the region of interest based on the degree of improvement;
A machine learning system characterized by having. Machine learning performed by a machine learning device that performs normal learning that generates a learning model for inferring a region of interest in input data based on pre-collected learning data, and transfer learning that updates the learning model. A method,
during the normal learning, generating the region of interest of the input data using the learning model, and performing a generation process of storing the generated region of interest in a pre-update region of interest storage unit as a pre-update region of interest; During or before the transfer learning, updating the region of interest of the input data from the pre-update region of interest to the post-update region of interest using the inputted person's knowledge, and storing the updated region of interest in the post-update region of interest storage unit. a step of performing the processing;
calculating a similarity distance indicating a change in the region of interest due to the update process;
During or before the transfer learning, dynamically selecting the input data to be subjected to the transfer learning based on the similarity distance;
A machine learning method characterized by having the following. A computer that performs normal learning that generates a learning model for inferring a region of interest of input data based on pre-collected learning data, and transfer learning that updates the learning model,
during the normal learning, generating the region of interest of the input data using the learning model, and performing a generation process of storing the generated region of interest in a pre-update region of interest storage unit as a pre-update region of interest; During or before the transfer learning, updating the region of interest of the input data from the pre-update region of interest to the post-update region of interest using the inputted person's knowledge, and storing the updated region of interest in the post-update region of interest storage unit. a step of processing;
calculating a similarity distance indicating a change in the region of interest due to the update process;
During or before the transfer learning, dynamically selecting the input data to be subjected to the transfer learning based on the similarity distance;
A machine learning program that runs.

Images