Computer system and model learning method

Abstract

A system manages a first model for solving one or more tasks, and a second model for generating playback input data in which input data constituting learning data used in learning of past tasks is reproduced. When new learning data for a new task has been accepted, the system generates replay learning data using the first model and the second model, and executes learning processing for updating the first model using the new learning data and the replay learning data. An index indicating uncertainty of the playback input data is calculated on the basis of an output obtained by inputting the playback input data into the updated first model, playback learning data for learning is selected on the basis of the index, and learning processing is executed using new learning data and the selected playback learning data.

Description

[0001] Cross-references

[0002] This application claims the priority of Japanese Patent Application No. 2022-151780 filed on September 22, 2022, the contents of which are incorporated herein by reference. Technical Field

[0003] The present invention relates to a technique for continuous learning for generating a model for solving multiple tasks. Background Art

[0004] In order to solve various tasks such as prediction and classification, systems and services that use models generated by machine learning have emerged. A learning method is known that uses an existing model to generate a model corresponding to a new task. However, in this learning method, it is known that the learning results of past tasks are lost as a destructive forgetting.

[0005] As a method of introducing learning results of past tasks and generating a model corresponding to a new task, the technology described in Non-Patent Document 1 is known.

[0006] Non-Patent Literature 1 describes continuous learning using a scalar including a generator that generates input data of a task learned in the past and a solver that solves the task learned in the past and a new task.

[0007] Prior art literature

[0008] Non-patent literature

[0009] Non-patent literature 1: Hanul Shin, Jung Kwon Lee, Jaehong Kim, Jiwon Kim, “Continual Learning with Deep Generative Replay”, May 24, 2017.

[0010] Non-patent literature 2: Yarin Gal, Zoubin Ghahramani, “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning”, June 6, 2015. Summary of the invention

[0011] Problems to be solved by the invention

[0012] The reliability of the data generated by the generator is not considered in Non-Patent Document 1. In the present invention, a system and a method are proposed that realize continuous learning that takes into account the reliability of the data generated by the generator.

[0013] Means for solving problems

[0014] A representative example of the invention disclosed in the present application is as follows. That is, a computer system having a computer, the computer having a processor, a storage device connected to the processor, and a connection interface connected to the processor, the computer system managing a first model for solving one or more tasks and a second model for generating replay input data, wherein the replay input data reproduces input data constituting learning data used in learning past tasks, the computer generating the replay input data using the second model when receiving new learning data related to a new task and consisting of new input data and new correct solution data, the computer generating replay learning data consisting of the replay input data and correct solution data generated by inputting the replay input data into the first model, the computer using the new learning data and the replay learning data to execute a method for converting the current first model into a new learning data. The model is updated to a first learning process of the first model for solving the new task and past tasks, the computer calculates an indicator for representing the uncertainty of the data input to the first model based on the output obtained by inputting the replay input data into the updated first model, the computer selects the replay learning data for learning based on the indicator of the replay input data, the computer uses the new learning data and the selected replay learning data to perform the first learning process, and the computer uses the new input data constituting the new learning data and the replay input data constituting the selected replay learning data to perform the following second learning process: updating the current second model to a second model that generates replay input data that reproduces the new input data and the selected replay input data.

[0015] Effects of the Invention

[0016] According to the present invention, continuous learning can be achieved taking into account the reliability of the data generated by the generator. As a result, the accuracy of the model can be improved. Through the description of the following embodiments, the topics, structures and effects other than the above become clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a diagram showing an example of the configuration of a computer according to the first embodiment.

[0018] Figure 2 This is a diagram for explaining the learning method of the model in the computer of Example 1.

[0019] Figure 3 This is a diagram showing the learning flow of the solver of the first embodiment.

[0020] Figure 4 This is a flowchart for explaining an example of the learning process of the solver in the computer of the first embodiment.

[0021] Figure 5 This is a diagram showing the learning process of the generator of Example 1.

[0022] Figure 6 This is a flowchart illustrating an example of the learning process of the generator in the computer of Example 1.

[0023] Figure 7 This is a diagram illustrating the learning method of the generator of Example 1.

[0024] Figure 8 This is a flowchart for explaining an example of the learning process of the solver in the computer of the second embodiment.

[0025] Fig. 9 This is a diagram showing an example of a screen displayed by a computer in the second embodiment. DETAILED DESCRIPTION

[0026] Hereinafter, embodiments of the present invention will be described using the accompanying drawings. However, the present invention is not limited to the description of the embodiments shown below. It is easy for a person skilled in the art to understand that the specific structure can be changed within the scope of the idea or purpose of the present invention.

[0027] In the configuration of the invention described below, the same or similar configurations or functions are denoted by the same reference numerals, and duplicate descriptions are omitted.

[0028] The terms “first”, “second”, “third”, etc. in this specification and the like are used to identify constituent elements, and do not necessarily limit the number or order.

[0029] In order to facilitate understanding of the invention, the position, size, shape, range, etc. of each structure shown in the drawings and the like sometimes do not represent the actual position, size, shape, range, etc. Therefore, the present invention is not limited to the position, size, shape, range, etc. disclosed in the drawings and the like.

[0030] Example 1

[0031] Figure 1 This is a diagram showing an example of the configuration of the computer 100 according to the first embodiment.

[0032] The computer 100 includes a processor 101, a memory 102, and a network interface 103. The hardware elements are connected to each other via an internal bus. The computer 100 may also include input devices such as a keyboard, a mouse, and a touch panel, and output devices such as a display.

[0033] The memory 102 stores programs executed by the processor 101 and information used by the programs. The memory 102 is also used as a work area for temporarily storing data.

[0034] The processor 101 executes a program stored in the memory 102. The processor 101 executes processing according to the program, thereby operating as a functional unit (module) for realizing a specific function. In the following description, when a process is described with a functional unit as the subject, it means that the processor 101 executes a program for realizing the functional unit.

[0035] The network interface 103 communicates with the outside via a network such as a WAN (Wide Area Network) and a LAN (Local Area Network).

[0036] The memory 102 of the first embodiment stores a program for realizing the task execution unit 110 and the learning unit 111. In addition, the memory 102 holds model management information 120.

[0037] The model management information 120 stores model information, which is used to manage the model of the solution task. The model information includes the structure and hyperparameters of the model.

[0038] The task execution unit 110 executes processing for solving one or more tasks using the model managed by the model management information 120. For example, the task execution unit 110 executes event prediction, data classification, etc. The present invention is not limited to the content of the executed tasks. In addition, it is not limited to the number of executed tasks.

[0039] For example, the task execution unit 110 outputs tissue state, presence of lesions, contrast, imaging angle, etc. based on the X-ray image. In this case, the output of each of the tissue state, presence of lesions, contrast, and imaging angle corresponds to one task.

[0040] The learning unit 111 performs a learning process for generating a model used by the task execution unit 110 .

[0041] Figure 2 This is a diagram for explaining the model learning method in the computer 100 of the first embodiment.

[0042] The learning section 111 performs learning using a scalar 200 composed of a generator 201 , a solver 202 , and an uncertainty index calculation unit 203 .

[0043] The generator 201 is a model that generates replay input data that reproduces the input data of all tasks learned so far. The solver 202 is a model that solves all tasks learned so far. The uncertainty index calculation unit 203 is a functional unit that calculates an index representing the uncertainty of the replay input data.

[0044] When the learning data of Task 1 is input, the learning unit 111 learns the generator 201 that generates the replay input data that reproduces the input data constituting the learning data of Task 1. Furthermore, the learning unit 111 learns the solver 202 that solves Task 1 using the learning data of Task 1.

[0045] When the learning data of task k (k is an integer greater than or equal to 2) is input, the learning unit 111 uses the scalar (k-1) obtained through the learning process of task (k-1) and the input data constituting the learning data to learn the generator 201 that generates the replay input data that reproduces the input data of tasks 1 to k. In addition, the learning unit 111 uses the learning data of task k and the learning data generated using the scalar (k-1) to learn the solver 202 that solves tasks 1 to k.

[0046] The model management information 120 stores a generator 201 of a scalar 200 generated by learning each task (see Figure 2 ) and solver 202 (refer to Figure 2 ) model information.

[0047] Figure 3 This is a diagram showing the learning process of the solver 202 of the first embodiment. Figure 4 This is a flowchart for explaining an example of the learning process of the solver 202 in the computer 100 according to the first embodiment.

[0048] Here, the scalar 200 obtained by the learning process so far is described as scalar (old) 200, and the scalar 200 of the new task is described as scalar (new) 200. The learning data 300 is the learning data of the new task, and is composed of input data (x) and correct solution data (y).

[0049] The learning unit 111 generates replay input data (x′) using the generator 201 of the scalar (old) 200 (step S401 ).

[0050] The learning unit 111 generates correct solution data (y') by inputting the replayed input data (x') to the solver 202 of the scalar (old) 200 (step S402).

[0051] The learning unit 111 learns the solver 202 of the scalar (new) 200 using the learning data 300 and the reproduced learning data 301 composed of the reproduced input data (x′) and the correct solution data (y′) (step S403 ).

[0052] The learning unit 111 calculates the uncertainty index of the replayed input data (x') (step S404). Specifically, the learning unit 111 inputs the replayed input data (x') constituting the replayed learning data 301 to the solver 202 of the scalar (new) 200. The learning unit 111 calculates the uncertainty index of the replayed input data (x') by inputting the output obtained from the solver 202 to the uncertainty index calculation unit 203.

[0053] The uncertainty of data in machine learning is also called Aleatoric Uncertainty. This indicator can be calculated using the Monte Carlo dropout method described in Non-Patent Document 2, for example. In the Monte Carlo dropout method, measures are performed to randomly make the weight of the model 0 for inference multiple times. In this way, the uncertainty of the inference result can be calculated. In addition, the uncertainty of the model and the uncertainty of the data can be quantified by calculating the histogram, mean, entropy or variance of the distribution of the results. In addition, the calculation method of the uncertainty of the data is not limited.

[0054] The learning unit 111 selects the replay learning data 301 to be used based on the uncertainty index of the replay input data (x') (step S405). For example, the learning unit 111 selects the replay learning data 301 consisting of the replay input data (x') whose index is less than a threshold value (low uncertainty). The threshold value is set in advance.

[0055] The learning unit 111 uses the learning data 300 and the selected replay learning data 301 to learn the solver 202 of the scalar (new) 200 (step S406). The solver 202 is learned using a known learning method. The learning method of the solver 202 is not limited.

[0056] Furthermore, the learning unit 111 may also calculate the uncertainty index of the input data (x) constituting the learning data 300. In the first embodiment, the learning data 300 used for learning is selected based on the uncertainty index of the input data (x), thereby improving the accuracy of the solver 202 and the generator 201.

[0057] Figure 5 This is a diagram showing the learning process of the generator 201 of Example 1. Figure 6 This is a flowchart for explaining an example of the learning process of the generator 201 in the computer 100 of the first embodiment. Figure 7This is a diagram illustrating the learning method of the generator 201 of Example 1.

[0058] After the learning process of the solver 202 is completed, the learning unit 111 starts the learning process of the generator 201.

[0059] The learning unit 111 generates replay input data (x′) using the generator 201 of the scalar (old) 200 (step S601 ).

[0060] The learning unit 111 selects the replayed input data (x') to be used based on the uncertainty index of the replayed input data (x') (step S602). The process of step S602 is performed using the process result of step S405.

[0061] The learning unit 111 learns the generator 201 using the input data (x) and the selected replayed input data (x′) (step S603 ).

[0062] Learn to use CGAN (Conditional Generative Adversarial Network). Figure 7 As shown, in CGAN, input data and conditional vectors (labels) are used as input to learn the discriminator and generator. Figure 7 The model corresponds to the solver 202 of Example 1.

[0063] For example, the Loss function shown in formula (1) is used to learn the generator 201. Here, D(x|y) represents the score when the real image and condition vector are input to the discriminator, and D(G(x|y)) represents the score when the image generated by the generator and the condition vector are input to the discriminator. σ represents the weight coefficient, and U represents the uncertainty index calculated by the uncertainty index calculation unit 203. z represents the potential variable of the generated image.

[0064] [Number 1]

[0065]

[0066] The first and second terms correspond to Loss 1, and the third term corresponds to Loss 2. As shown in equation (1), the first embodiment is characterized in that a term that takes into account the uncertainty index calculated based on the output of the solver 202 is added.

[0067] According to Embodiment 1, by selecting the replayed input data (x') used in the learning of the generator 201 based on the uncertainty of the replayed input data (x'), the accuracy of the replayed input data (x') generated by the generator 201 can be improved. In addition, similarly, by selecting the replayed learning data 301 used in the learning of the solver 202, the accuracy of the solver 202 can be improved.

[0068] Furthermore, the task execution unit 110 and the learning unit 111 may be realized by using a computer system composed of a plurality of computers 100. In addition, the model management information 120 may be stored in an external system.

[0069] Example 2

[0070] The computer 100 of the second embodiment receives the correction of the replay input data (x') and the correct solution data (y') constituting the replay learning data 301, and learns the generator 201 and the solver 202. The second embodiment will be described below, focusing on the differences from the first embodiment.

[0071] The hardware structure and software structure of the computer 100 of the second embodiment are the same as those of the first embodiment.

[0072] In the second embodiment, the learning method of the solver 202 is partially different. Figure 8 This is a flowchart for explaining an example of the learning process of the solver 202 in the computer 100 according to the second embodiment. Fig. 9 This is a diagram showing an example of a screen displayed by the computer 100 according to the second embodiment.

[0073] The processing from step S401 to step S404 of the second embodiment is the same as that of the first embodiment.

[0074] After executing the process of step S404 , the learning unit 111 displays the screen 900 (step S451 ) and waits for the user's operation.

[0075] Screen 900 includes an indicator column 901 , an input data column 902 , a correct solution data column 903 , a delete button 904 , an input data correction button 905 , a correct solution data correction button 906 , and a learning execution button 907 .

[0076] Indicator column 901 displays the uncertainty indicator of the replayed input data (x'). Fig. 9 In the index column 901, a graph is displayed in which the horizontal axis represents the uncertainty index and the vertical axis represents the probability of the prediction result. One point corresponds to one replay input data (x'). The user selects the replay input data (x') to be referenced from the index column 901.

[0077] The input data column 902 displays the replayed input data (x'). The correct solution data column 903 displays the correct solution data (y') paired with the replayed input data (x').

[0078] The delete button 904 is an operation button for deleting the replay learning data 301 composed of the replay input data (x′) from the data set.

[0079] The input data correction button 905 is an operation button for correcting the playback input data (x'). The data can be corrected by directly operating the input data field 902 or executing a preset correction process.

[0080] Correct solution data correction button 906 is an operation button for correcting the correct solution data (y') paired with the replay input data (x'). The correct solution data column 903 may be directly operated to correct the data, or a preset correction process may be executed.

[0081] The learning execution button 907 is an operation button for instructing to execute learning of the solver 202 again using the replayed learning data 301 selected by the user.

[0082] When receiving the user's operation (step S452), the learning unit 111 determines whether it is an operation to delete the replayed learning data 301 (step S453).

[0083] If it is a deletion operation of the replay learning data 301 , the learning unit 111 deletes the designated replay learning data 301 (step S454 ), and then shifts to a waiting state.

[0084] If the operation is not a deletion operation of reproducing the learning data 301 , the learning unit 111 determines whether the operation is a correction operation of reproducing any one of the input data (x′) and the correct solution data (y′) (step S455 ).

[0085] In the case of a correction operation of either the replay input data (x') or the correct solution data (y'), the learning unit 111 corrects the data according to the correction operation (step S456), and then shifts to a waiting state.

[0086] When receiving the instruction to execute learning, the learning unit 111 executes the processing of step S405 and step S406. The processing of step S405 and step S406 in the second embodiment is the same as that in the first embodiment.

[0087] The learning process of the generator 201 of the second embodiment is the same as that of the first embodiment. However, the learning is performed using the corrected playback learning data 301 .

[0088] The user can delete or modify the replay learning data (x') by referring to the uncertainty index, etc. Thus, a high-precision model can be generated.

[0089] Furthermore, the learning data 300 can also be corrected or deleted using the screen 900 .

[0090] In addition, the present invention is not limited to the above-mentioned embodiments, and includes various modified examples. In addition, for example, the above-mentioned embodiments are examples of detailed structures for easy understanding of the present invention, and are not limited to having all the structures described. In addition, with respect to a part of the structure of each embodiment, other structures can be added, deleted, or replaced.

[0091] In addition, the above-mentioned structures, functions, processing parts, processing units, etc. can also be designed by integrated circuits, etc., so that part or all of them can be realized by hardware. In addition, the present invention can also be realized by the program code of the software for realizing the functions of the embodiments. In this case, a storage medium with program code recorded is provided to the computer, and the processor possessed by the computer reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-mentioned embodiments, and the program code itself and the storage medium storing the program code constitute the present invention. As a storage medium for supplying such a program code, for example, a floppy disk, a CD-ROM, a DVD-ROM, a hard disk, an SSD (Solid State Drive), an optical disk, a magneto-optical disk, a CD-R, a magnetic tape, a non-volatile memory card, a ROM, etc. are used.

[0092] In addition, the program code that realizes the functions described in this embodiment can be implemented by a wide range of programs or script languages ​​such as assembler, C / C++, perl, Shell, PHP, Python, Java, etc.

[0093] Furthermore, the program code of the software used to implement the functions of the embodiment can also be distributed via a network, thereby saving it to a storage unit such as a computer's hard disk or memory or a storage medium such as a CD-RW or CD-R, and the processor of the computer reads and executes the program code stored in the storage unit or the storage medium.

[0094] In the above-mentioned embodiments, the control lines and information lines are those considered necessary for the description, and not all control lines and information lines are necessarily shown in the product. All structures may be connected to each other.

Claims

1. A computer system comprising a computer, the computer comprising a processor, a storage device connected to the processor, and a connection interface connected to the processor, It is characterized in that The computer system manages a first model that solves one or more tasks and a second model that generates replay input data that reproduces input data that constitutes learning data used in learning past tasks, The computer generates the replay input data using the second model when receiving new learning data consisting of new input data and new correct solution data related to the new task, the computer generates replay learning data consisting of the replay input data and correct solution data generated by inputting the replay input data into the first model, The computer performs a first learning process for updating the current first model to the first model that solves the new task and the past tasks using the new learning data and the replayed learning data, The computer calculates an index representing uncertainty of data input to the first model based on an output obtained by inputting the replayed input data to the updated first model, the computer selecting the replay learning data for learning according to the index of the replay input data, the computer performs the first learning process using the new learning data and the selected replayed learning data, The computer uses the new input data constituting the new learning data and the replay input data constituting the selected replay learning data to perform the following second learning process: updating the current second model to generate the second model of the replay input data that reproduces the new input data and the selected replay input data.

2. The computer system according to claim 1, characterized in that: After calculating the index of the replay input data, the computer generates display information for displaying the replay learning data and the index of the replay input data. The computer receives at least one of an instruction to modify and an instruction to delete the playback learning data via a screen displayed according to the display information.

3. The computer system according to claim 1, characterized in that: When calculating the index of the replayed input data, the computer calculates the index of the new input data based on the output obtained by inputting the new input data into the updated first model, The computer selects the new learning data for learning according to the indicator of the new input data, the computer performs the first learning process using the selected new learning data and the selected replayed learning data, The computer performs the second learning process using the new input data constituting the selected new learning data and the replayed input data constituting the selected replayed learning data.

4. A method for learning a model for solving more than one task, performed by a computer system, characterized in that: The computer system includes a computer having a processor, a storage device connected to the processor, and a connection interface connected to the processor. The computer system manages a first model that solves one or more tasks and a second model that generates replay input data that reproduces input data that constitutes learning data used in learning past tasks, The learning method of the model includes: In a first step, the computer generates the replay input data using the second model when receiving new learning data consisting of new input data and new correct solution data related to the new task; In a second step, the computer generates replay learning data consisting of the replay input data and correct solution data generated by inputting the replay input data into the first model; In a third step, the computer uses the new learning data and the replayed learning data to perform a first learning process for updating the current first model to the first model for solving the new task and the past task; In a fourth step, the computer calculates an index representing uncertainty of data input to the first model based on an output obtained by inputting the replayed input data to the updated first model; In a fifth step, the computer selects the replay learning data for learning according to the index of the replay input data; In a sixth step, the computer uses the new learning data and the selected replayed learning data to perform the first learning process; and In the seventh step, the computer uses the new input data constituting the new learning data and the replay input data constituting the selected replay learning data to perform the following second learning process: updating the current second model to generate the second model of the replay input data that reproduces the new input data and the selected replay input data.

5. The model learning method according to claim 4, characterized in that: The fourth step comprises the following steps: After calculating the index of the replay input data, the computer generates display information for displaying the replay learning data and the index of the replay input data; as well as The computer receives at least one of an instruction to modify and an instruction to delete the playback learning data via a screen displayed according to the display information.

6. The model learning method according to claim 4, characterized in that: The fourth step comprises the following steps: the computer calculates the index of the new input data according to the output obtained by inputting the new input data into the updated first model, The fifth step includes the following steps: the computer selects the new learning data for learning according to the index of the new input data, The sixth step includes the following steps: the computer uses the selected new learning data and the selected replayed learning data to perform the first learning process, The seventh step includes the following steps: the computer uses the new input data constituting the selected new learning data and the replayed input data constituting the selected replayed learning data to perform the second learning process.

Images