Model training method, model training program, and information processing device
By inputting multiple pseudo-data into the trained mechanical learning model, calculating its credibility and determining data with a reference value less than that, it is used to train a new model, and solving the inefficiency improvement of the inference resistance of the mechanical learning model members in the prior art. Problem, achieving efficient member reasoning patience generation.
Patent Information
- Application Number
- CN202280094107.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-31
- Publication Date
- 2024-11-12
AI Technical Summary
The prior art is inefficient in improving the member reasoning tolerance of mechanical learning models, and it is difficult to efficiently generate models with member reasoning tolerance.
By inputting multiple pseudo-data into the trained first category classification model, the credibility of the positive solution label is obtained, and the pseudo-data with the reference value less than is determined as training data based on the confidence and distance calculation, the new category classification model is trained. To enhance member reasoning patience.
It realizes efficient generation of mechanical learning models with member inference tolerance, reduces pseudo-data affecting member inference tolerance, and improves training efficiency.
Smart Images

Figure CN118946897A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a model training method, a model training program and an information processing device. Background Art
[0002] In recent years, the development and use of systems using machine learning have been rapidly advancing. On the other hand, inherent security issues have also been found in systems using machine learning. For example, member inference attacks are known as one of the security issues.
[0003] In a membership inference attack, for example, it is inferred whether the data that the attacker is interested in is included in the training data of the machine learning model of the target of the attack.
[0004] As a defense method against membership inference attacks, a method of training a machine learning model using pseudo data as training data is known. The pseudo data may be generated by adding noise to basic data or may be generated by machine learning based on basic data.
[0005] In addition, there is a known technique for filtering training data. In one example, there is a known training data filtering device that filters training data that can shorten training time (see Patent Document 1).
[0006] The property that it is not easy to infer whether certain data is included in the training data is sometimes called resistance to membership inference attacks. Various pseudo data may include data that affects resistance to membership inference attacks (hereinafter referred to as "membership inference resistance").
[0007] In the past, the training of machine learning models was repeatedly performed by dividing pseudo data into several groups, and the reasoning tolerance of each member was observed to evaluate each group. The group division method was changed and this process was repeated several times, and data commonly used in models with low tolerance was determined as data that reduced the reasoning tolerance of members and was excluded from the training data.
[0008] Patent document 1: Japanese Patent Application Publication No. 2021-107970;
[0009] Patent Document 1: Japanese Patent Application Publication No. 2019-159961;
[0010] Patent document 3: U.S. Patent Application Publication No. 2007 / 0143284.
[0011] However, the determination of data that reduces the reasoning resistance of members by such conventional methods requires a long period of calculation, and therefore there is a problem that it is difficult to efficiently train a machine learning model to improve the reasoning resistance of members. Summary of the invention
[0012] In one aspect, the present invention aims to efficiently generate a machine learning model that is resistant to member reasoning.
[0013] Therefore, in the model training method, the following processing is performed by a computer: by inputting multiple processed data into a first category classification model, the credibility of the correct answer label is obtained for each of the multiple processed data, wherein the first category classification model is a model trained using basic data corresponding to the correct answer label, and the multiple processed data are respectively corresponding to the correct answer label and are respectively different from the basic data; determining the processed data corresponding to the credibility less than a first reference value; and using the determined processed data as training data to train a new category classification model.
[0014] According to one embodiment, a machine learning model with member reasoning resistance can be efficiently generated. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a diagram illustrating a hardware configuration of an information processing device as an example of an embodiment.
[0016] Figure 2 This is a diagram illustrating a functional configuration of an information processing device as an example of an embodiment.
[0017] Figure 3 This is a diagram for explaining an example of a second training execution unit and a second credibility vector acquisition unit in an information processing device as an example of an embodiment.
[0018] Figure 4 This is a diagram for explaining the processing contents in the information processing device as an example of the embodiment.
[0019] Figure 5 This is a diagram showing an overview of the training phase and the inference phase of a category classification model in an information processing device as an example of an implementation.
[0020] Figure 6 This is a flowchart for explaining a model training method in an information processing device as an example of an implementation method.
[0021] Figure 7 This is a flowchart for explaining a first example of a method for determining dummy data in an information processing device as an example of an embodiment.
[0022] Figure 8 This is a flowchart for explaining a second example of the method for determining dummy data in the information processing device as an example of the embodiment. DETAILED DESCRIPTION
[0023] Hereinafter, the implementation methods of the model training method, model training program, and information processing device will be described with reference to the accompanying drawings. However, the implementation methods shown below are merely illustrative, and are not intended to exclude the application of various modifications and technologies not explicitly stated in the implementation methods. That is, the present implementation method can be implemented by various modifications without departing from its main purpose. In addition, each figure does not only have the content of the structural elements shown in the figure, but can also include other functions, etc.
[0024] (A) Structure
[0025] Figure 1 This is a diagram illustrating a hardware configuration of an information processing device 1 as an example of the embodiment.
[0026] For example Figure 1 As shown, the information processing device 1 includes a processor 11, a memory 12, a storage device 13, a graphics processing device 14, an input interface 15, an optical drive device 16, a device connection interface 17, and a network interface 18 as components. These components 11 to 18 are configured to be able to communicate with each other via a bus 19. The information processing device 1 is an example of a computer.
[0027] The processor 11 (control unit) controls the entire information processing device 1. The processor 11 may also be a multiprocessor. The processor 11 may be, for example, a CPU, an MPU (Micro Processing Unit), a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), an FPGA (Field Programmable Gate Array), or a GPU (Graphics Processing Unit). In addition, the processor 11 may also be a combination of two or more elements of a CPU, an MPU, a DSP, an ASIC, a PLD, an FPGA, or a GPU.
[0028] The processor 11 executes the control program (model training program 13a) to achieve Figure 2 The function of the training processing unit 100 is illustrated in FIG.
[0029] The information processing device 1 realizes the function as the training processing unit 100 by executing, for example, a model training program 13 a or an OS (Operating System) program recorded in a computer-readable non-transitory recording medium.
[0030] The program describing the processing content executed by the information processing device 1 can be pre-recorded in various recording media. For example, the model training program 13a executed by the information processing device 1 can be pre-stored in the storage device 13. The processor 11 loads at least a portion of the model training program 13a in the storage device 13 into the memory 12 and executes the loaded model training program 13a.
[0031] In addition, the model training program 13a for executing the information processing device 1 (processor 11) may be pre-recorded on a non-temporary removable recording medium such as an optical disk 16a, a memory device 17a, or a memory card 17c. The model training program 13a stored on the removable recording medium can be executed after being installed in the storage device 13 under the control of the processor 11. In addition, the processor 11 may directly read the model training program 13a from the removable recording medium and execute it.
[0032] The memory 12 is a storage memory including a ROM (Read Only Memory) and a RAM (Random Access Memory). The RAM of the memory 12 is used as a main storage device of the information processing device 1. The RAM temporarily stores at least a part of the OS program and the control program executed by the processor 11. In addition, the memory 12 stores various data required for the processing by the processor 11.
[0033] The storage device 13 is a storage device such as a hard disk drive (HDD), an SSD (Solid State Drive), or a storage class memory (SCM), and stores various data. The storage device 13 is used as an auxiliary storage device of the information processing device 1. The storage device 13 stores an OS program, a control program, and various data. The control program includes a model training program 13a.
[0034] As the auxiliary storage device, a semiconductor storage device such as SCM or flash memory may be used. In addition, a plurality of storage devices 13 may be used to form a RAID (Redundant Arrays of Inexpensive Disks).
[0035] In addition, the storage device 13 may store various data acquired or generated by the third training execution unit 101 , the basic data acquisition unit 102 , the dummy data acquisition unit 103 , the first training execution unit 104 , the second training execution unit 105 , and the final training data generation unit 106 described later.
[0036] A monitor 14a is connected to the graphics processing device 14. The graphics processing device 14 displays an image on the screen of the monitor 14a according to a command from the processor 11. Examples of the monitor 14a include a display device using a CRT (Cathode Ray Tube) and a liquid crystal display device.
[0037] The input interface 15 is connected to a keyboard 15a and a mouse 15b. The input interface 15 sends signals sent from the keyboard 15a and the mouse 15b to the processor 11. In addition, the mouse 15b is an example of a pointing device, and other pointing devices can also be used. As other pointing devices, a touch panel, a tablet, a touch pad, a trackball, etc. are listed.
[0038] The optical drive device 16 uses laser light or the like to read data recorded on the optical disk 16a. The optical disk 16a is a removable non-temporary recording medium that records data that can be read by reflection of light. Examples of the optical disk 16a include DVD (Digital Versatile Disc), DVD-RAM, CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), and the like.
[0039] The device connection interface 17 is a communication interface for connecting a peripheral device to the information processing device 1. For example, a memory device 17a and a memory reader / writer 17b can be connected to the device connection interface 17. The memory device 17a is a non-temporary recording medium equipped with a communication function with the device connection interface 17, such as a USB (Universal Serial Bus) memory. The memory reader / writer 17b writes data to the memory card 17c or reads data from the memory card 17c. The memory card 17c is a card-type non-temporary recording medium.
[0040] The network interface 18 is connected to a network (not shown). The network interface 18 may be connected to other information processing devices and communication devices via the network. For example, data related to a disease may be input via the network.
[0041] Figure 2 1 is a diagram illustrating a functional configuration of an information processing device 1 as an example of an embodiment. Figure 2 As shown, the information processing device 1 has a function as a training processing unit 100 .
[0042] In the information processing device 1 , the processor 11 executes the control program (model training program 13 a ), thereby realizing the function as the training processing unit 100 .
[0043] The training processing unit 100 implements a learning process (training process) in machine learning using training data. That is, the information processing device 1 functions as a training device that trains a machine learning model through the training processing unit 100.
[0044] The training processing unit 100 includes a third training execution unit 101, which uses the training data (teacher data) to which the correct answer label is assigned to realize the learning process in the machine learning. In this example, the training processing unit 100 includes a data screening unit 100a that screens (determines) the training data input to the third training execution unit 101. The "correct answer label" can be the correct answer information assigned to each data.
[0045] The training data input to the third training execution unit 101 may be a plurality of pseudo data generated by adding noise to original data for defense against member inference attacks. “Pseudo data” is an example of processed data obtained by processing original data.
[0046] The data filter 100a excludes data that affects the reasoning tolerance of members, that is, data that reduces the reasoning tolerance of members from a plurality of pseudo data. The data filter 100a filters the training data used for the new classification model (third classification model C) trained by the third training execution unit 101 .
[0047] The category classification model is a machine learning model for classifying data into multiple categories. The machine learning model may also be a deep learning model (deep neural network), for example. The neural network may be a hardware circuit or a software-based virtual network that connects layers virtually constructed on a computer program by the processor 11 or the like.
[0048] You can Figure 2 As shown, the data screening unit 100 a includes a basic data acquisition unit 102 , a dummy data acquisition unit 103 , a first training execution unit 104 , a second training execution unit 105 , and a finalized training data generation unit 106 .
[0049] The basic data acquisition unit 102 acquires basic data. The basic data is data (teacher data) associated with a correct answer label. The basic data is training data used by the first training execution unit 104 to implement a learning process in machine learning.
[0050] The basic data may be data generated (processed) based on the collected unprocessed raw data, or may be the raw data itself. However, compared to the basic data being the raw data itself, it is preferred that the basic data be data processed based on the raw data. In the case where the raw data is data with a confidentiality level above a predetermined level, such as disease-related data, it is preferred not to use the raw data itself for training as much as possible from the perspective of maintaining confidentiality. However, depending on the content of the data, the raw data may also be used as the basic data.
[0051] The basic data acquisition unit 102 may acquire basic data generated by an external device, or may generate basic data within the information processing device 1 .
[0052] The pseudo data acquisition unit 103 acquires a plurality of pseudo data. The pseudo data acquisition unit 103 may generate pseudo data based on original data.
[0053] The pseudo data is an example of processed data generated (processed) based on the collected unprocessed original data. The pseudo data acquisition unit 103 can use various known methods to generate pseudo data. For example, the pseudo data can be generated by applying noise to the original data. In one example, the pseudo data acquisition unit 103 can generate each pseudo data by applying random noise to the original data. The noise can be Gaussian noise or Laplace noise. In one example, the pseudo data can be data obtained by processing the basic data.
[0054] Alternatively, the pseudo data acquisition unit 103 may train a generative model based on machine learning such as GAN (Generative Adversarial Network) using the original data, and generate pseudo data using the trained model. Alternatively, the pseudo data acquisition unit 103 may generate pseudo data using dynamic programming (DP).
[0055] The processing degree of each pseudo data may be greater than that of the basic data. The processing degree refers to the degree of processing from the original data. In one example, the greater the noise added to the original data, the greater the processing degree.
[0056] The plurality of dummy data are associated with the correct answer labels, respectively. However, the plurality of dummy data are different from the basic data, respectively. The plurality of dummy data include training data (teacher data) used by the second training execution unit 105 to implement the learning process in machine learning.
[0057] The pseudo data acquisition unit 103 may acquire pseudo data generated by a device outside the information processing device 1, or may generate pseudo data within the information processing device 1. In particular, the pseudo data acquisition unit 103 may generate a plurality of pseudo data within the information processing device 1 based on the basic data acquired by the basic data acquisition unit 102.
[0058] The first training execution unit 104 uses the basic data as training data to perform a first class classification model A (model A: refer to Figure 4 ) is trained to generate a trained first category classification model A. The first category classification model A is an example of a first category classification model. Basic data is, for example, a combination of input data x and correct output data y. Preferably, the first training execution unit 104 uses multiple basic data to perform training of the first category classification model A. The first training execution unit 104 can perform training of the first category classification model A using a known method.
[0059] The training of the first classification model A performed by the first training execution unit 104 using the basic data is sometimes referred to as first training. In addition, the classification model before training by the first training execution unit 104 may be an empty machine learning model. The machine learning model may also be simply referred to as a model.
[0060] The second training execution unit 105 uses a plurality of pseudo data as training data to train the second category classification model B (model B) and generates a trained second category classification model B. The second category classification model is an example of a second category classification model. The plurality of pseudo data are, for example, respectively constituted as a combination of input data x and correct output data y. The second training execution unit 105 can perform the training of the second category classification model B using a known method.
[0061] The second training execution unit 105 may use two or more first dummy data (for example, dummy data #1 described later: see Figure 3 ) as training data to perform the second classification model B (for example, model B1: refer to Figure 3 ) training.
[0062] The training of the second classification model B performed by the second training execution unit 105 using a plurality of dummy data is sometimes referred to as second training. In addition, the classification model before training by the second training execution unit 105 may be an empty machine learning model that is the same as the classification model before training by the first training execution unit 104.
[0063] In addition, the second training execution unit 105 may also use dummy data as training data to train multiple (eg, two) second category classification models B (eg, models B1 and B2) to generate multiple trained second category classification models B.
[0064] Figure 3 This is a diagram for explaining an example of processing by the second training execution unit 105 and the second credibility vector acquisition unit 108 in the information processing device 1 as an example of the embodiment.
[0065] In this Figure 3 In the example shown, the second training execution unit 105 performs training of two second class classification models B1 and B2. The second training execution unit 105 may include an allocation unit 111. The allocation unit 111 allocates the plurality of dummy data acquired from the dummy data acquisition unit 103 into a plurality of groups. The allocation unit 111 may randomly allocate the plurality of dummy data into a plurality of groups.
[0066] exist Figure 3 In the example shown, the dummy data is allocated to dummy data #1 (first conversion data) belonging to one group and dummy data #2 (second conversion data) belonging to another group different from the one group. However, the dummy data may be allocated to more than three groups. The dummy data #1 and the dummy data #2 are two or more dummy data, respectively.
[0067] The second training execution unit 105 trains the second class classification model B1 using the dummy data #1. The second training execution unit 105 trains the second class classification model B2 using the dummy data #2.
[0068] Figure 2 The determined training data generating unit 106 determines (selects and generates) the training data used by the third training execution unit 101 to implement the learning process in the machine learning. The determined training data generating unit 106 can exclude data that may deteriorate the member's reasoning tolerance from the plurality of pseudo data. The determined training data generating unit 106 can determine the training data that maintains the member's reasoning tolerance from the plurality of pseudo data.
[0069] The determination training data generating unit 106 may screen the training data of the third training executing unit 101 using the trained first category classification model A, the trained second category classification model B, and a plurality of pseudo data as evaluation objects.
[0070] In addition, the training data generation unit 106 may also obtain the trained first category classification model A, the trained second category classification model B, and a plurality of pseudo data as evaluation objects from outside the information processing device 1. In this case, the device outside the information processing device 1 may have the functions of the basic data acquisition unit 102, the pseudo data acquisition unit 103, the first training execution unit 104, and the second training execution unit 105.
[0071] like Figure 2 As shown, the determination training data generating unit 106 includes a first credibility vector acquiring unit 107 , a second credibility vector acquiring unit 108 , a distance calculating unit 109 , and a determining unit 110 .
[0072] The first credibility vector acquisition unit 107 acquires a first credibility vector VA for each of the plurality of pseudo data by inputting the plurality of pseudo data to the first classification model A. Generation of the first credibility vector VA is one of the inference processes using the trained first classification model A, and is referred to as first inference.
[0073] The credibility vector includes the credibility of each label, which is the result of the classification of data by the classification model, as an element. "Label" can be an item for classifying data by the classification model. The credibility is the probability that the combination of the data of interest and the label (item) is correct.
[0074] In one example, when the category classification model classifies the input data into four elements, such as element (A), element (B), element (C), and element (D), the credibility is calculated for each label. In addition, the credibility of the correct answer label of the input data is calculated. The credibility vector includes the credibility of each label as an element.
[0075] The first credibility vector acquisition unit 107 is an example of a credibility acquisition unit that inputs a plurality of pseudo data into the first class classification model A, thereby acquiring the credibility of the correct answer label for each of the plurality of pseudo data.
[0076] The second credibility vector acquisition unit 108 acquires a second credibility vector VB including the credibility of each of the plurality of labels as a result of discrimination for each of the two or more second processed data. Generation of the second credibility vector VB is one of the inference processes using the trained second class classification model B, and is called second inference.
[0077] The second credibility vector acquisition unit 108 inputs pseudo data to the second class classification model B to perform inference, and acquires a second credibility vector VB.
[0078] When the second training execution unit 105 performs training of a plurality of second class classification models B, the second credibility vector acquisition unit 108 inputs pseudo data to each of the plurality of second class classification models B to perform inference, and acquires a second credibility vector VB.
[0079] The second credibility vector acquisition unit 108 may include a switching unit 112. The switching unit 112 replaces (exchanges) the pseudo data input to each of the second class classification models B1 and B2 in the training phase and the evaluation phase.
[0080] exist Figure 3In the example shown, the switching unit 112 inputs the pseudo data #2 as the pseudo data for evaluation to the second category classification model B1 trained using the pseudo data #1. On the other hand, the switching unit 112 inputs the pseudo data #1 as the pseudo data for evaluation to the second category classification model B2 trained using the pseudo data #2.
[0081] In this way, the switching unit 112 can avoid the situation where the second category classification model B1 trained with the pseudo data #1 evaluates the same pseudo data #1 as in the training stage by replacing the pseudo data input to the second category classification model B1 and B2 in the training stage and the evaluation stage. Similarly, it can avoid the situation where the second category classification model B2 trained with the pseudo data #2 evaluates the same pseudo data #2 as in the training stage.
[0082] When the same data is used in the training phase and the evaluation phase, the credibility of the correct answer label in the second credibility vector VB may be higher than necessary due to the influence of overlearning, and the distance |VA-VB| may not reflect the reasoning tolerance of the members. Figure 3 The structure makes it easy to evaluate the reasoning tolerance of members based on the distance |VA-VB|.
[0083] The switching unit 112 can evaluate the member inference tolerance for the entirety of the pseudo data #1 and the pseudo data #2 by replacing (exchanging) the pseudo data input to each of the second class classification models B1 and B2.
[0084] The second credibility vector acquisition unit 108 may use two or more first dummy data (for example, dummy data #1, #2: see Figure 3 ) after training the second category classification model B1, B2 (refer to Figure 3 ).
[0085] The second credibility vector acquisition unit 108 obtains two or more second dummy data (for example, dummy data #1 and #2 described later: see Figure 3 ) is input into the trained second category classification model B2, B1 (refer to Figure 3 ), and generate a second credibility vector VB (refer to Figure 4 ). The first dummy data (eg, dummy data #1) and the second dummy data (eg, dummy data #2) may be different.
[0086] Figure 2 The distance calculation unit 109 shown acquires the distance between the first credibility vector VA and the second credibility vector VB.
[0087] The distance can be the KL distance (Kulbeck-Leibler distance) or the L1 distance (also called the Manhattan distance). In one example, the first credibility vector VA is set to VA(p1, ... pn) (where p1, ... pn are the credibility of each label in the first credibility vector VA). The second credibility vector VB is set to VB(q1, ... qn) (where q1, ... qn are the credibility of each label in the second credibility vector VB).
[0088] The KL distance |VA-VB| between the first credibility vector VA and the second credibility vector VB is given by the following equation (1).
[0089] sqrt((p1-q1) 2 +···+(pn-qn) 2 )···(1)
[0090] The L1 distance |VA-VB| between the first credibility vector VA and the second credibility vector VB is given by the following equation (2).
[0091] |p1-q1|+···+|pn-qn|···(2)
[0092] The determination unit 110 determines the pseudo data to be input to the third training execution unit 101 from among the plurality of pseudo data. The determination unit 110 may determine the pseudo data based on the first credibility vector VA.
[0093] In one example, the determination unit 110 may determine pseudo data based on the first credibility vector VA. The determination unit 110 determines whether the credibility of the correct answer label of the first credibility vector VA is less than a first reference value. The determination unit 110 may determine pseudo data corresponding to the credibility less than the first reference value as data that does not have a negative impact on the member's reasoning tolerance. Pseudo data determined as data that does not have a negative impact on the member's reasoning tolerance in this way may be used as training data for training the third category classification model C (model C) by the third training execution unit 101. The first reference value may be a predetermined threshold.
[0094] The determination unit 110 may further determine pseudo data based on the distance |VA-VB| between the first credibility vector VA and the second credibility vector VB. For example, the determination unit 110 determines whether the distance |VA-VB| is a value greater than the second reference value. Even if it is pseudo data corresponding to a credibility greater than the first reference value, the determination unit 110 may determine pseudo data whose distance |VA-VB| is greater than the second reference value as data that does not have a negative impact on the member's reasoning tolerance. The determination unit 110 may use pseudo data such as data determined as data that does not have a negative impact on the member's reasoning tolerance as training data for training the third category classification model C. That is, the determination unit 110 may determine pseudo data that satisfies the conditions that the credibility of the correct answer label of the first credibility vector VA is less than the first reference value, or the distance |VA-VB| is a value greater than the second reference value, as training data for the third category classification model C.
[0095] In other words, in this embodiment, the determination unit 110 excludes pseudo data whose credibility of the correct answer label of the first credibility vector VA is greater than the first reference value and whose distance |VA-VB| is less than the second reference value from the plurality of pseudo data. Pseudo data that increases the credibility of the correct answer label of the first credibility vector VA and decreases the distance |VA-VB| may worsen the influence on the member's reasoning tolerance. Therefore, the determination unit 110 can prevent pseudo data that may worsen the influence on the member's reasoning tolerance from occurring.
[0096] However, the determination unit 110 may determine the corresponding pseudo data as training data for the third category classification model C when the credibility of the correct answer label of the first credibility vector VA is less than the first reference value and the distance |VA-VB| is greater than the second reference value.
[0097] The third training execution unit 101 uses the pseudo data determined by the determination unit 110 as training data to train the third category classification model C. The third category classification model C is a model actually used for inference. The pseudo data determined by the determination unit 110 is composed of, for example, a combination of input data x and correct output data y.
[0098] The training of the third category classification model C performed by the third training execution unit 101 using a plurality of determined pseudo data is sometimes referred to as third training. In addition, the category classification model before training by the third training execution unit 101 may be an empty machine learning model that is the same as the category classification model of the previous stage trained by the first training execution unit 104 or the second training execution unit 105.
[0099] Figure 4This is a diagram for explaining the processing contents of the first training execution unit 104 , the second training execution unit 105 , the first credibility vector acquisition unit 107 , and the second credibility vector acquisition unit 108 in the information processing device 1 as an example of the embodiment.
[0100] In this Figure 4 , an example is shown in which the second training execution unit 105 performs training on two second class classification models B1 and B2 , and the second credibility vector acquisition unit 108 acquires a second credibility vector VB using these multiple second class classification models B1 and B2 .
[0101] In this Figure 4 In the example shown, the processing content is roughly divided into processing #1 and processing #2. Processing #1 is the training stage of the first category classification model A and the second category classification models B1 and B2. Processing #2 is the evaluation stage of pseudo data #1 and #2. The evaluation stage is an example of the inference stage based on the first category classification model A and the second category classification models B1 and B2.
[0102] Process #1 includes first training and second training. In the first training, the first training execution unit 104 uses the basic data as training data to perform training of the first category classification model A. In the second training, the second training execution unit 105 uses the dummy data #1 as training data to perform training of the second category classification model B1. In the second training, the second training execution unit 105 further uses the dummy data #2 as training data to perform training of the second category classification model B2.
[0103] Process #2 includes the first inference and the second inference. In the first inference, the first credibility vector acquisition unit 107 inputs the pseudo data (both pseudo data #1 and pseudo data #2) into the trained first class classification model A to acquire the first credibility vector VA.
[0104] In the second inference, the second credibility vector acquisition unit 108 inputs the pseudo data #1 to the trained second class classification model B2. As a result, the second credibility vector acquisition unit 108 acquires a second credibility vector VB for the pseudo data #1.
[0105] In the second inference, the second credibility vector acquisition unit 108 inputs the pseudo data #2 to the trained second class classification model B1. As a result, the second credibility vector acquisition unit 108 acquires a second credibility vector VB for the pseudo data #2.
[0106] In this way, the second credibility vector acquisition unit 108 can acquire the second credibility vector VB for the dummy data (both dummy data #1 and dummy data #2).
[0107] The determination unit 110 excludes pseudo data in which the credibility of the correct answer label of the first credibility vector VA is greater than the first reference value and the distance |VA-VB| between the first credibility vector VA and the second credibility vector VB is less than the second reference value. The determination unit 110 can determine pseudo data that affects the member's reasoning tolerance based on the credibility of the correct answer label and the distance |VA-VB|. In other words, the determination unit 110 determines pseudo data, which is used as training data for training the third category classification model C by the third training execution unit 101.
[0108] Figure 5 This is a diagram showing an overview of the training phase and the inference phase of a category classification model in the information processing device 1 as an example of an embodiment.
[0109] Figure 5 The process shown includes a training phase. The training phase involves passing Figure 4 The third training is to use the pseudo data determined by the process shown as training data to train the third class classification model C. The third class classification model C is a new class classification model and is actually used in the inference stage.
[0110] The third training execution unit 101 trains an empty machine learning model using the determined pseudo data as training data, thereby setting parameters of the machine learning model.
[0111] In the inference stage, if query data x to be the object of category classification is input to the third category classification model C, the third category classification model C outputs the category classification result as output data y.
[0112] In the information processing device 1 of this embodiment, in the inference stage, for example, by inputting disease function data or the like as query data x into the third category classification model C, it can be applied as a device for inferring whether a certain disease is suspected. However, the information processing device 1 is not limited to this case, and can also be applied as a device for inferring whether an email article is junk or not, and other various category classification devices.
[0113] (B) Action
[0114] according to Figure 6 The flowchart (steps S1 to S5 ) shown here explains a method of training a category classification model (machine learning model) in the information processing device 1 as an example of the embodiment configured as described above.
[0115] In step S1, the dummy data acquisition unit 103 generates a plurality of dummy data. The dummy data acquisition unit 103 can generate a plurality of dummy data based on the original data. Information constituting the dummy data is stored in a predetermined storage area such as the storage device 13.
[0116] In step S2 , the first training execution unit 104 executes the first training to train the first category classification model A using the basic data as training data. The empty category classification model before the first training may be pre-stored in the storage device 13 . The trained first category classification model A may be stored in the storage device 13 .
[0117] In step S3 , the second training execution unit 105 executes second training for training the second class classification model B ( B1 , B2 ) using a plurality of dummy data as training data. The empty class classification model before the second training is executed may be stored in advance in the storage device 13 .
[0118] In step S4 , the data filtering unit 100 a identifies (filters) pseudo data to be input to the third training execution unit 101 in order to train the third class classification model C.
[0119] Specifically, the first credibility vector acquisition unit 107 generates a first credibility vector VA for each of the plurality of pseudo data by inputting the plurality of pseudo data into the first category classification model A. The first credibility vector VA may include the credibility of each of the plurality of labels, which is a result of discrimination, as an element. In particular, the first credibility vector VA includes the credibility of the correct label.
[0120] Similarly, the second credibility vector 10 generates a second credibility vector VB for each of the plurality of pseudo data by inputting the plurality of pseudo data into the second class classification model B.
[0121] The determination unit 110 determines pseudo data based on at least one of the reliability of the correct label inferred by the first class classification model A and the distance |VA-VB|.
[0122] In step S5, the third training execution unit 101 executes the third training for training the third class classification model C using the pseudo data determined in step S4 as training data. The third class classification model C trained in this way has member inference resistance.
[0123] Furthermore, the pseudo data acquisition unit 103 may also acquire a plurality of pseudo data generated by a device external to the information processing device 1. Furthermore, the information processing device 1 may acquire a first category classification model A and a second category classification model B generated by a device external to the information processing device 1. In these cases, the processing of steps S1, S2, and S3 may be omitted.
[0124] Figure 7 This is a flowchart for explaining a first example of a method for determining dummy data in the information processing device 1 as an example of the embodiment (steps S11 to S18 ). Figure 7 The flowchart shown is Figure 6 This is an example of the processing of step S4.
[0125] In step S11 , a first category classification model A trained using only basic data and a second category classification model B trained using only dummy data are prepared.
[0126] The data screening unit 100a determines whether there is any unevaluated dummy data remaining (step S12). If the result of the determination is that there is no unevaluated dummy data remaining (refer to the "No" path of step S12), the dummy data determination process ends. If there is any unevaluated dummy data remaining (refer to the "Yes" path of step S12), the process proceeds to step S13.
[0127] In step S13, the first credibility vector acquisition unit 107 selects an unevaluated pseudo data from the plurality of pseudo data. The first credibility vector acquisition unit 107 obtains a first credibility vector VA by inputting the selected pseudo data into the first category classification model A for inference. The first credibility vector VA includes the credibility of the correct answer label.
[0128] In step S14, the second credibility vector acquisition unit 108 obtains the second credibility vector VB by inputting the pseudo data selected in step S13 into the second category classification model B for inference. Figure 4 As shown, a plurality of second category classification models B1, B2 are prepared, and in the case of training and evaluation, the second credibility vector VB is obtained by replacing (exchanging) the pseudo data input to each second category classification model B1, B2.
[0129] In step S15, the determination unit 110 determines whether the credibility of the correct answer label in the first credibility vector VA is greater than the first reference value. If the credibility of the correct answer label is greater than the first reference value (refer to the "yes" path of step S15), the process proceeds to step S16. On the other hand, if the credibility of the correct answer label is less than the first reference value (refer to the "no" path of step S15), the process proceeds to step S17.
[0130] In step S16, the determination unit 110 determines whether the distance |VA-VB| between the first credibility vector VA and the second credibility vector VB is less than the second reference value. If the distance |VA-VB| is a value greater than the second reference value (refer to the "No" path of step S16), the process proceeds to step S17. If the distance |VA-VB| is less than the second reference value (refer to the "Yes" path of step S16), the process proceeds to step S18.
[0131] In step S17, the determination unit 110 determines the pseudo data as training data for the third category classification model C, and the process returns to step S12. Therefore, when the credibility of the correct answer label of the first credibility vector VA is less than the first reference, and even if the credibility is greater than the first reference, the distance |VA-VB| is greater than the second reference value, the pseudo data is determined as training data.
[0132] In step S18, the determination unit 110 excludes the pseudo data from the training data of the third category classification model C, and the process returns to step S12. Therefore, when the credibility of the correct answer label of the first credibility vector VA is greater than the first reference value and the distance |VA-VB| is less than the second reference value, the pseudo data is excluded from the training data.
[0133] Figure 8 This is a flowchart for explaining a second example of the method for determining dummy data in the information processing device 1 as an example of the embodiment (steps S21 to S28 ). Figure 8 The flowchart shown is Figure 6 Another example of the processing of step S4.
[0134] exist Figure 8 In the process of steps S21 to S24, Figure 7 The processing of steps S11 to S14 in is the same, and the description of each processing is omitted.
[0135] In step S25, the determination unit 110 determines whether the credibility of the correct answer label in the first credibility vector VA is greater than the first reference value. If the credibility of the correct answer label is less than the first reference value (refer to the "No" path of step S25), the process proceeds to step S26. On the other hand, if the credibility of the correct answer label is greater than the first reference value (refer to the "Yes" path of step S25), the process proceeds to step S28.
[0136] In step S26, the determination unit 110 determines whether the distance |VA-VB| between the first credibility vector VA and the second credibility vector VB is less than the second reference value. If the distance |VA-VB| is a value greater than the second reference value (refer to the "No" path of step S26), the process proceeds to step S27. If the distance |VA-VB| is less than the second reference value (refer to the "Yes" path of step S26), the process proceeds to step S28.
[0137] In step S27, the determination unit 110 determines the pseudo data as training data for the third category classification model C, and the process returns to step S22. Therefore, when the credibility of the correct answer label of the first credibility vector VA is less than the first reference, and the distance |VA-VB| is a value greater than the second reference value, the pseudo data is determined as training data.
[0138] In step S28, the determination unit 110 excludes the pseudo data from the training data of the third category classification model C, and the process returns to step S22. Therefore, when the credibility of the correct answer label of the first credibility vector VA is greater than the first reference value, or the distance |VA-VB| is less than the second reference value, it is excluded from the training data.
[0139] Information constituting each of the first class classification model A, the second class classification model B, and the third class classification model C is stored in a predetermined storage area such as the storage device 13 .
[0140] (C) Effects of One Implementation Method
[0141] In one embodiment of the method, a computer performs the following processing, inputs a plurality of pseudo data to a first category classification model A trained using basic data corresponding to a correct answer label, and obtains the credibility of the correct answer label for each of the plurality of pseudo data. Furthermore, the computer performs the following processing to determine pseudo data corresponding to a credibility less than a first reference value. The computer performs the following processing to train a new category classification model, i.e., a third category classification model C, using the determined pseudo data as training data.
[0142] According to the above method, the third class classification model C is trained by excluding the pseudo data that affects the member reasoning resistance. In this way, a machine learning model having the member reasoning resistance can be generated.
[0143] In addition, the basic data and the processed data are generated based on the collected unprocessed raw data. The processed data is more processed from the raw data than the basic data. Therefore, a plurality of pseudo data are respectively input to the first category classification model A trained with the basic data closer to the raw data, so that pseudo data that affects the reasoning tolerance of the members can be effectively excluded.
[0144] In addition, the information processing device 1 performs the following processing, by inputting a plurality of pseudo data into the first category classification model A, and generating a first credibility vector VA with the credibility of each of the discrimination results, i.e., a plurality of labels, as an element for each of the plurality of pseudo data. The information processing device 1 inputs two or more pseudo data #2 (second pseudo data) different from the pseudo data #1 into the second category classification model B1 trained using two or more pseudo data #1 (first pseudo data) among the plurality of pseudo data. Thus, the information processing device 1 generates a second credibility vector VB with the credibility of each of the discrimination results, i.e., a plurality of labels, as an element for each of the two or more second processed data. Moreover, the information processing device 1 performs the processing of obtaining the distance |VA-VB| between the first credibility vector VA and the second credibility vector VB. The information processing device 1 performs the following processing to determine pseudo data in the pseudo data, in which the credibility of the correct answer label of the first credibility vector VA is greater than the first reference value and the distance is greater than the second reference value.
[0145] According to the above method, even when the reliability of the correct answer label of the first reliability vector VA is greater than the first reference value due to some noise, etc., a detailed check can be performed based on the distance |VA-VB| to determine the pseudo data as training data.
[0146] Alternatively, pseudo data whose credibility of the correct answer label of the first credibility vector VA is less than the first reference value and whose distance |VA-VB| is greater than the second reference value may be determined as training data. In this case, pseudo data that affects the reasoning tolerance of members can be excluded.
[0147] (D) Others
[0148] The disclosed technology is not limited to the above-mentioned embodiment, and can be implemented with various modifications within the scope of the present embodiment. For example, each structure and each process of the present embodiment can be selected or discarded as needed, or can be appropriately combined.
[0149] For example, in the above-mentioned embodiment, an example of using two second category classification models B1 and B2 using dummy data as training data is shown, but the present invention is not limited to this, and three or more second category classification models B may be used.
[0150] In this case, the switching unit 112 performs control so that the pseudo data input to each second class classification model B is different in the training phase and the evaluation phase.
[0151] Description of Reference Numerals
[0152] 1…information processing device; 11…processor; 12…memory; 13…storage device; 13a…model training program; 14…graphics processing device; 14a…monitor; 15…input interface; 15a…keyboard; 15b…mouse; 16…optical drive device; 16a…optical disk; 17…machine connection interface; 17a…storage device; 17b…storage reader / writer; 17c…storage card; 18…network interface; 19…bus; 100…training processing unit; 100a…data screening unit; 101…third training execution unit; 102…basic data acquisition unit; 103…pseudo data acquisition unit; 104…first training execution unit; 105…second training execution unit; 106…determination training data generation unit; 107…first credibility vector acquisition unit; 108…second credibility vector acquisition unit; 109…distance calculation unit; 110…determination unit; 111…allocation unit; 112…switching unit.
Claims
1. A model training method, wherein: The following processing is performed by a computer: By inputting a plurality of processed data into a first category classification model, the credibility of the correct answer label is respectively obtained for the plurality of processed data, wherein the first category classification model is a model trained using basic data corresponding to the correct answer label, and the plurality of processed data are respectively corresponding to the correct answer label and are respectively different from the basic data, determining processed data corresponding to the reliability less than a first reference value, The determined processed data is used as training data to train a new category classification model.
2. The model training method according to claim 1, wherein: The basic data and the processed data are generated based on the collected unprocessed raw data. The processed data is processed to a greater extent from the original data compared to the basic data.
3. The model training method according to claim 1 or 2, wherein: The computer performs the following processing: By inputting the plurality of processed data into the first category classification model, a first credibility vector is generated for each of the plurality of processed data, which has the credibility of each of the plurality of labels as a result of discrimination as an element. By inputting two or more second processed data into a second category classification model, second credibility vectors are generated for the two or more second processed data, each of which has a credibility of a discrimination result, i.e., a plurality of labels, as an element. The second category classification model is a model trained using two or more first processed data among the plurality of processed data, and the second processed data is data among the plurality of processed data that is different from the first processed data. obtaining a distance between the first credibility vector and the second credibility vector, The second processed data whose credibility of the correct answer label of the first credibility vector among the two or more second processed data is greater than the first reference value and whose distance is greater than the second reference value is also determined as the training data for training the new category classification model. The new category classification model is trained using the determined training data.
4. The model training method according to claim 1 or 2, wherein: The computer performs the following processing: By inputting the plurality of processed data into the first category classification model, a first credibility vector is generated for each of the plurality of processed data, which has the credibility of each of the plurality of labels as a result of discrimination as an element. By inputting two or more second processed data into the second category classification model, second credibility vectors are generated for the two or more second processed data, each of which has a credibility of a discrimination result, i.e., a plurality of labels, as an element, wherein the second category classification model input is a model trained using two or more first processed data among the plurality of processed data, and the second processed data is data among the plurality of processed data that is different from the first processed data, obtaining a distance between the first credibility vector and the second credibility vector, determining the second processed data in which the reliability of the correct answer label of the first reliability vector is less than the first reference value and the distance is greater than a second reference value, among the two or more second processed data, The new category classification model is trained using the determined second processed data as the training data.
5. A model training procedure, wherein: The computer is made to execute the following processing: By inputting a plurality of processed data into a first category classification model, the credibility of the correct answer label is respectively obtained for the plurality of processed data, wherein the first category classification model is a model trained using basic data corresponding to the correct answer label, and the plurality of processed data are respectively corresponding to the correct answer label and are respectively different from the basic data, determining processed data corresponding to the reliability less than a first reference value, The determined processed data is used as training data to train a new category classification model.
6. The model training program according to claim 5, wherein: The computer is caused to execute the following processing: The basic data and the processed data are generated based on the collected unprocessed raw data. The processed data is processed to a greater extent from the original data compared to the basic data.
7. The model training program according to claim 5 or 6, wherein: By inputting the plurality of processed data into the first category classification model, a first credibility vector is generated for each of the plurality of processed data, which has the credibility of each of the plurality of labels as a result of discrimination as an element. By inputting two or more second processed data into a second category classification model, second credibility vectors are generated for the two or more second processed data, each of which has a credibility of a discrimination result, i.e., a plurality of labels, as an element. The second category classification model is a model trained using two or more first processed data among the plurality of processed data, and the second processed data is data among the plurality of processed data that is different from the first processed data. obtaining a distance between the first credibility vector and the second credibility vector, The second processed data whose credibility of the correct answer label of the first credibility vector among the two or more second processed data is greater than the first reference value and whose distance is greater than the second reference value is also determined as the training data for training the new category classification model. The new category classification model is trained using the determined training data.
8. The model training program according to claim 5 or 6, wherein: The computer is caused to execute the following processing: By inputting the plurality of processed data into the first category classification model, a first credibility vector is generated for each of the plurality of processed data, which has the credibility of each of the plurality of labels as a result of discrimination as an element. By inputting two or more second processed data into a second category classification model, second credibility vectors are generated for the two or more second processed data, each of which has a credibility of a discrimination result, i.e., a plurality of labels, as an element. The second category classification model is a model trained using two or more first processed data among the plurality of processed data, and the second processed data is data among the plurality of processed data that is different from the first processed data. obtaining a distance between the first credibility vector and the second credibility vector, determining the second processed data in which the reliability of the correct answer label of the first reliability vector is less than the first reference value and the distance is greater than a second reference value, among the two or more second processed data, The new category classification model is trained using the determined second processed data as the training data.
9. An information processing device, characterized in that: A control unit is provided, which performs the following processing: By inputting a plurality of processed data into a first category classification model, the credibility of the correct answer label is respectively obtained for the plurality of processed data, wherein the first category classification model is a model trained using basic data corresponding to the correct answer label, and the plurality of processed data are respectively corresponding to the correct answer label and are respectively different from the basic data, determining processed data corresponding to the reliability less than a first reference value, The determined processed data is used as training data to train a new category classification model.
10. The information processing device according to claim 9, wherein: The basic data and the processed data are generated based on the collected unprocessed raw data. The processed data is processed to a greater extent from the original data compared to the basic data.
11. The information processing device according to claim 9 or 10, wherein: The control unit executes the following processing: By inputting the plurality of processed data into the first category classification model, a first credibility vector is generated for each of the plurality of processed data, which has the credibility of each of the plurality of labels as a result of discrimination as an element. By inputting two or more second processed data into a second category classification model, second credibility vectors are generated for the two or more second processed data, each of which has a credibility of a discrimination result, i.e., a plurality of labels, as an element. The second category classification model is a model trained using two or more first processed data among the plurality of processed data, and the second processed data is data among the plurality of processed data that is different from the first processed data. obtaining a distance between the first credibility vector and the second credibility vector, The second processed data whose credibility of the correct answer label of the first credibility vector among the two or more second processed data is greater than the first reference value and whose distance is greater than the second reference value is also determined as the training data for training the new category classification model. The new category classification model is trained using the determined training data.
12. The information processing device according to claim 9 or 10, wherein: The control unit executes the following processing: By inputting the plurality of processed data into the first category classification model, a first credibility vector is generated for each of the plurality of processed data, which has the credibility of each of the plurality of labels as a result of discrimination as an element. By inputting two or more second processed data into a second category classification model, second credibility vectors are generated for the two or more second processed data, each of which has a credibility of a discrimination result, i.e., a plurality of labels, as an element. The second category classification model is a model trained using two or more first processed data among the plurality of processed data, and the second processed data is data among the plurality of processed data that is different from the first processed data. obtaining a distance between the first credibility vector and the second credibility vector, determining the second processed data in which the reliability of the correct answer label of the first reliability vector is less than the first reference value and the distance is greater than a second reference value, among the two or more second processed data, The new category classification model is trained using the determined second processed data as the training data.
