Training data generation device and training data generation method

Abstract

A training data generation device comprising: a secondary training dataset generation unit that generates a secondary training dataset from a primary training dataset; and a tertiary training data generation unit that generates a tertiary training dataset from the secondary training dataset. The secondary training dataset generation unit includes: a probability distribution generation unit that generates, from the primary training dataset, a latent data space corresponding to the primary training dataset, and generates a probability distribution in the latent space; a gain function generation unit that generates a gain function used for extracting primary input data expected to have a large effect on improvement of model performance through use of the latent data space and the probability distribution; and a synthesis data generation unit that extracts primary input data as the source for generation of synthesis data on the basis of a score of the gain function and generates the secondary training dataset having secondary input data generated from the extracted primary input data as an element.

Description

Training data generation device and training data generation method

[0001] The present invention relates to a technique for generating training data used in supervised learning.

[0002] Currently, machine learning, also known as supervised learning, is used in a variety of application fields. As the range of application fields expands, there are also increasing opportunities to handle more sensitive data as training data, which is the data used for learning. It is also becoming common for training data to be continuously updated along with the model. Here, training data in supervised learning refers to a set of input data that serves as input for the model and data (hereinafter referred to as labels) that indicate the evaluation results of the model's output for the input data. An example of a label is data that indicates whether the output is correct or incorrect.

[0003] As the application fields of supervised learning expand, there is a demand for techniques to conceal training data to prevent data leakage. For example, a method for concealing training data is described in Non-Patent Document 1. In the active synthetic data generation method described in Non-Patent Document 1, pseudo-data (hereinafter referred to as synthetic data) is generated based on actual data (hereinafter referred to as raw data) in order to conceal the training data. Then, after generating synthetic data from the raw data, an annotator annotates (hereinafter referred to as labeling) the synthetic data to generate training data. Using the training data generated by the active synthetic data generation method described in Non-Patent Document 1, it is possible to train a model whose inference performance is comparable to that of a model trained using raw data.

[0004] Osamu Zeisho, Takayuki Miura, Kazuki Iwahana, Masanori Kii, "Active Synthetic Data Generation: Utilizing Privacy-Preserving Synthetic Data for Machine Learning with Annotation," Information Processing Society of Japan SIG Technical Report, Vol. 2023-DPS-195, No. 9, pp. 1-8, May 2023.

[0005] However, the method described in Non-Patent Document 1 generates synthetic data from all raw data, which results in high processing costs in terms of the time and data volume required for generation.

[0006] Therefore, an object of the present invention is to provide an active synthesis data generation technique that can reduce the processing costs associated with generating training data.

[0007] In one aspect of the present invention, a primary training data set is a set whose elements are primary training data consisting of a pair of primary input data, which is raw data to be input to a model, and a primary label, which is raw data indicating an evaluation result of an output of the model for the primary input data; a secondary training data set generation unit that generates, from the primary training data set, a secondary training data set whose elements are secondary input data, which is synthetic data to be input to a model generated from the primary input data extracted by a predetermined method; and a tertiary training data generation unit that generates, from the secondary training data set, a tertiary training data set whose elements are tertiary training data consisting of a pair of tertiary input data, which is synthetic data to be input to a model generated by a predetermined method using the secondary training data set, and a tertiary label, which is raw data indicating an evaluation result of an output of the model for the tertiary input data. a training data generation unit and a secondary training data generation unit, wherein the secondary training data set generation unit includes: a probability distribution generation unit that generates a latent data space corresponding to a primary input data set from a primary input data set whose elements are primary input data that constitute primary training data, which are elements of the primary training data set, and generates a probability distribution in the latent data space; a gain function generation unit that generates a gain function used to extract primary input data that is expected to have great utility in improving model performance, using the latent data space and the probability distribution; and a synthetic data generation unit that extracts primary input data for which synthetic data is to be generated based on the score of the gain function, and generates the secondary training data set whose elements are secondary input data generated from the extracted primary input data.

[0008] In one aspect of the present invention, a primary training data set is a set whose elements are primary training data consisting of a pair of primary input data, which is raw data to be input to a model, and a primary label, which is raw data indicating an evaluation result of an output of the model for the primary input data; a secondary training data set generation unit that generates, from the primary training data set, a secondary training data set whose elements are secondary training data consisting of a pair of secondary input data, which is synthetic data to be input to a model generated from a pair of primary input data extracted by a predetermined method and a primary label corresponding to the primary input data, and a pair of secondary labels, which is synthetic data indicating an evaluation result of an output of the model for the secondary input data; and a tertiary training data set generation unit that generates, from the primary training data set, a secondary training data set whose elements are secondary training data consisting of a pair of tertiary input data, which is synthetic data to be input to a model generated by a predetermined method using the secondary training data set, and a pair of tertiary labels, which is raw data indicating an evaluation result of an output of the model for the tertiary input data. a tertiary training data generation unit for generating a tertiary training data set having training data as elements, wherein the secondary training data set generation unit includes: a probability distribution generation unit that generates a latent data space corresponding to a primary input data set from a primary input data set having as elements primary input data constituting primary training data, which is an element of the primary training data set, and generates a probability distribution in the latent data space; a gain function generation unit that generates a gain function using the latent data space and the probability distribution to extract primary input data that is expected to have great utility in improving model performance; and a synthetic data generation unit that extracts primary input data to be used for synthetic data generation based on the score of the gain function, and generates the secondary training data set having as elements secondary training data generated from a set of the extracted primary input data and a primary label corresponding to the primary input data.

[0009] One aspect of the present invention is a secondary training data set generation unit that defines a primary training data set as a set whose elements are primary input data, which is raw data that serves as an input to a model, and generates a secondary training data set from the primary training data set, whose elements are secondary input data, which is synthetic data that serves as an input to a model generated from the primary input data extracted by a predetermined method; and a tertiary training data generation unit that generates a tertiary training data set whose elements are tertiary training data composed of a set of tertiary input data, which is synthetic data that serves as an input to a model generated by a predetermined method using the secondary training data set, and tertiary labels, which are raw data that indicate an evaluation result of the output of the model for the tertiary input data. the secondary training data set generation unit includes: a probability distribution generation unit that generates a latent data space corresponding to the primary training data set from the primary training data set, and generates a probability distribution in the latent data space; a gain function generation unit that generates a gain function used to extract primary input data that is expected to have great utility for improving model performance, using the latent data space and the probability distribution; and a synthetic data generation unit that extracts primary input data for which synthetic data is to be generated based on the score of the gain function, and generates the secondary training data set whose elements are secondary input data generated from the extracted primary input data.

[0010] According to the present invention, since synthetic data is not generated from all raw data, it is possible to reduce the processing costs involved in generating training data.

[0011] Fig. 1 is a block diagram showing the configuration of a model learning device 100. Fig. 2 is a flowchart showing the operation of the model learning device 100. Fig. 3 is a block diagram showing the configuration of a secondary training data set generation unit 110. Fig. 4 is a flowchart showing the operation of the secondary training data set generation unit 110. Fig. 5 is a diagram showing an example of the functional configuration of a computer that realizes each device in an embodiment of the present invention.

[0012] Hereinafter, embodiments of the present invention will be described in detail. Components having the same functions are given the same numbers, and duplicated explanations will be omitted.

[0013] First Embodiment A model learning device 100 learns a model through supervised learning. Here, the model is a function, and is configured as, for example, a neural network.

[0014] The model learning device 100 will be described below with reference to FIGS. 1 and 2. FIG. 1 is a block diagram showing the configuration of the model learning device 100. FIG. 2 is a flowchart showing the operation of the model learning device 100. As shown in FIG. 1, the model learning device 100 includes a secondary training data set generation unit 110, a tertiary training data set generation unit 120, a model learning unit 130, and a recording unit 190. The recording unit 190 is a component that appropriately records information necessary for the processing of the model learning device 100. The recording unit 190 pre-records the model to be used at the start of learning. If the model is configured as a neural network, the recording unit 190 records the initial values ​​of the weights and biases of the neural network. The recording unit 190 may also pre-record the primary training data set instead of using it as input to the model learning device 100. Here, the primary training data set is a set whose elements are primary training data, each of which is a set of primary input data, which is raw data used as input to the model, and primary labels, which are raw data indicating the evaluation results of the model's output for the primary input data.

[0015] The operation of the model learning device 100 will be described with reference to FIG.

[0016] In S110, the secondary training data set generation unit 110 receives the primary training data set as input, and generates and outputs a secondary training data set whose elements are secondary input data, which is synthetic data that will be input to a model generated from the primary input data extracted by a predetermined method. Instead of generating a secondary training data set whose elements are secondary input data, the secondary training data set generation unit 110 may generate a secondary training data set whose elements are secondary training data consisting of a pair of secondary input data, which is synthetic data that will be input to a model generated from a pair of primary input data extracted by a predetermined method and primary labels corresponding to the primary input data, and secondary labels, which is synthetic data that indicates an evaluation result of the model output for the secondary input data.

[0017] An example of the secondary training data set generation unit 110 will be described below with reference to Figures 3 and 4. Figure 3 is a block diagram showing the configuration of the secondary training data set generation unit 110. Figure 4 is a flowchart showing the operation of the secondary training data set generation unit 110. As shown in Figure 3, the secondary training data set generation unit 110 includes a probability distribution generation unit 111, a gain function generation unit 112, and a synthetic data generation unit 113.

[0018] The operation of the secondary training data set generator 110 will now be described with reference to FIG.

[0019] In S111, the probability distribution generation unit 111 receives a primary training data set as input, generates a latent data space corresponding to the primary input data set from the primary input data set whose elements are primary input data constituting the primary training data, which are elements of the primary training data set, generates a probability distribution in the latent data space, and outputs the latent data space and the probability distribution. Here, the latent data of the primary input data refers to data obtained by reducing the dimensions of the primary input data, and the latent data space corresponding to the primary input data set refers to a set whose elements are the latent data of the primary input data. The probability distribution generation unit 111 can generate the probability distribution using, for example, a Bayesian network or a generative adversarial network (GAN).

[0020] In S112, the gain function generation unit 112 receives the latent data space and probability distribution generated in S111 as input, and generates and outputs a gain function used to extract primary input data expected to have a large utility for improving model performance using the latent data space and probability distribution generated in S111. The gain function generation unit 112 may generate the gain function using the model currently being trained in addition to the latent data space and probability distribution generated in S111. The gain function generation unit 112 can generate a gain function used to extract primary input data expected to have a large utility for improving model performance using, for example, the probability distribution generated in S111 and a gain function BALD (see Reference Non-Patent Document 1) based on uncertainty in the latent data space generated in S111.

[0021] (Reference Non-Patent Document 1: Yarin Gal, Riashat Islam, Zoubin Ghahramani, “Deep Bayesian Active Learning with Image Data,” [online], [Retrieved June 3, 2024], Internet<URL: https: / / arxiv.org / abs / 1703.02910> In S113, the synthetic data generation unit 113 receives the gain function generated in S112 as an input, extracts primary input data for which synthetic data is to be generated based on the score of the gain function, and generates and outputs a secondary training data set whose elements are secondary input data generated from the extracted primary input data by a predetermined synthetic data generation method.Instead of generating a secondary training data set whose elements are secondary input data, the synthetic data generation unit 113 may extract primary input data for which synthetic data is to be generated based on the score of the gain function generated in S112, and generate a secondary training data set whose elements are secondary training data generated by a predetermined synthetic data generation method from a set of the extracted primary input data and a primary label corresponding to the primary input data.

[0022] The synthetic data generation unit 113 may extract data that is likely to produce a high score through, for example, an experimental design using Bayesian optimization as primary input data for generating synthetic data. The predetermined synthetic data generation method may be a method of generating synthetic data after adding predetermined Gaussian noise to generation parameters in order to guarantee differential privacy.

[0023] In S120, the tertiary training data generation unit 120 receives the secondary training data set generated in S110 as input, generates a tertiary training data set whose elements are tertiary input data, which is synthetic data that serves as input to a model generated using the secondary training data set by a predetermined method, and outputs the tertiary training data. The tertiary training data set is composed of tertiary input data, which is synthetic data that serves as input to a model generated using the secondary training data set by a predetermined method, and tertiary labels, which are raw data indicating the evaluation results of the model's output for the tertiary input data. Here, the predetermined method may be a method of generating raw data indicating the evaluation results of the model's output for the secondary input data, and defining the pair of the secondary input data and the raw data as a pair of tertiary input data and tertiary labels, if the secondary training data set is a set whose elements are secondary training data, using the secondary labels that constitute the secondary training data, and defining the pair of the secondary input data and the raw data as a pair of tertiary input data and tertiary labels. Here, the raw data may be generated manually, for example. Alternatively, the predetermined method may be a method of generating raw data indicating the evaluation results of the model's output for the secondary input data that constitute the secondary training data, using the secondary labels that constitute the secondary training data, and defining the pair of the secondary input data and the raw data as a pair of tertiary input data and tertiary labels, if the secondary training data set is a set whose elements are secondary training data. Here, the raw data may be manually generated by referring to, for example, secondary labels.

[0024] The number of elements contained in the tertiary training data set may be any number, and may be the same as the number of elements contained in the secondary training data set, for example.

[0025] In S130, the model learning unit 130 receives the tertiary training data set generated in S120 as input, learns a model using the tertiary training data set, and outputs the model. Here, any supervised learning method can be used as the learning method.

[0026] According to an embodiment of the present invention, since synthetic data is not generated for all raw data, it is possible to reduce the processing costs associated with generating training data. Furthermore, since tertiary labels are generated as raw data, the labels do not include the influence of noise that occurs when generating synthetic data in a manner that guarantees differential privacy, and therefore training can be performed while achieving a strict privacy budget and maintaining model performance.

[0027] Second Embodiment In the first embodiment, the primary training data set is a set whose elements are primary training data, which is a set of primary input data and primary labels. However, the primary training data set may also be a set whose elements are primary input data. A model learning device 100 according to the second embodiment will be described below with reference to FIGS. 1 and 2. As shown in FIG. 1, the configuration of the model learning device 100 is the same as that of the model learning device 100 according to the first embodiment. The operation of the model learning device 100 will be described below with reference to FIG. 2.

[0028] In S110, the secondary training data set generation unit 110 takes the primary training data set as input, generates and outputs a secondary training data set whose elements are secondary input data, which is synthetic data that will be input to a model generated from primary input data extracted using a predetermined method.

[0029] An example of the secondary training data set generation unit 110 will be described with reference to Figures 3 and 4. The configuration of the secondary training data set generation unit 110 is the same as the configuration of the secondary training data set generation unit 110 in the first embodiment, as shown in Figure 3. The operation of the secondary training data set generation unit 110 will be described below with reference to Figure 4.

[0030] In S111, the probability distribution generation unit 111 receives a primary training data set as input, generates a latent data space corresponding to the primary training data set from the primary training data set, generates a probability distribution in the latent data space, and outputs the latent data space and the probability distribution.

[0031] In S112, the gain function generation unit 112 receives the latent data space and probability distribution generated in S111 as input, and generates and outputs a gain function used to extract primary input data that is expected to have a great effect on improving the performance of the model, using the latent data space and probability distribution generated in S111.

[0032] In S113, the synthetic data generation unit 113 receives the gain function generated in S112 as input, extracts primary input data for which synthetic data is to be generated based on the score of the gain function, and generates and outputs a secondary training data set whose elements are secondary input data generated from the extracted primary input data using a predetermined synthetic data generation method.

[0033] In S120, the tertiary training data generation unit 120 receives the secondary training data set generated in S110 as input, generates a tertiary training data set whose elements are tertiary input data, which is synthetic data that serves as input to a model generated using the secondary training data set by a predetermined method, and tertiary training data consisting of a pair of tertiary labels, which are raw data indicating the evaluation results of the model's output for the tertiary input data. Here, the predetermined method may be a method of generating raw data indicating the evaluation results of the model's output for the secondary input data, and treating the pair of the secondary input data and the raw data as a pair of tertiary input data and tertiary labels. Here, the raw data may be generated manually, for example.

[0034] In S130, the model learning unit 130 receives the tertiary training data set generated in S120 as input, learns a model using the tertiary training data set, and outputs the model.

[0035] According to an embodiment of the present invention, since synthetic data is not generated for all raw data, it is possible to reduce the processing costs associated with generating training data. Furthermore, since tertiary labels are generated as raw data, the labels do not include the influence of noise that occurs when generating synthetic data in a manner that guarantees differential privacy, and therefore training can be performed while achieving a strict privacy budget and maintaining model performance.

[0036] Third Embodiment The training data generating configuration included in the model learning device 100 of the first and second embodiments can also be made into an independent device. If this device is referred to as a training data generation device 200, the training data generation device 200 will include a secondary training data set generation unit 110, a tertiary training data set generation unit 120, and a recording unit 190 (see FIG. 1). The operation of the training data generation device 200 is the same as S110 and S120 in the operation of the model learning device 100 (see FIG. 2).

[0037] According to the embodiment of the present invention, since synthetic data is not generated from all raw data, it is possible to reduce the processing costs involved in generating training data.

[0038] <Additional Notes> The functions performed by the components described herein may be implemented in circuitry or processing circuitry, including general-purpose processors, application-specific processors, integrated circuits, ASICs (Application Specific Integrated Circuits), CPUs (Central Processing Units), conventional circuits, and / or combinations thereof, programmed to perform the described functions. A processor includes transistors and other circuits and is considered to be circuitry or processing circuitry. A processor may also be a programmed processor that executes a program stored in a memory.

[0039] In this specification, a circuitry, unit, or means is hardware that is programmed to realize or performs the described functions, which may be any hardware disclosed herein or any hardware known to be programmed to realize or perform the described functions.

[0040] If the hardware is a processor considered to be a type of circuitry, the circuitry, means, or unit is a combination of the hardware and software used to configure the hardware and / or processor.

[0041] The various processes described above can be implemented by loading a program that executes each step of the above method into the recording unit 2020 of the computer 2000 shown in Figure 5, and operating the control unit 2010, input unit 2030, output unit 2040, display unit 2050, etc.

[0042] The program describing the processing contents can be recorded on a computer-readable recording medium, which may be, for example, a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, or any other suitable recording medium.

[0043] The program may be distributed by, for example, selling, transferring, lending, etc. portable recording media such as DVDs and CD-ROMs on which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and then transferred from the server computer to other computers via a network, thereby distributing the program.

[0044] A computer that executes such a program may first temporarily store the program recorded on a portable recording medium or transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored in its storage device and executes the process in accordance with the read program. Alternatively, the computer may read the program directly from a portable recording medium and execute the process in accordance with the program. Furthermore, the computer may execute the process in accordance with the program each time a program is transferred from a server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may instead execute the process through a so-called ASP (Application Service Provider) service, which realizes the processing function by issuing an execution instruction and obtaining the results. Furthermore, the server computer may execute the process at the terminal using a so-called SaaS (Software as a Service) service, which allows users to use part of the server computer along with the program. In this embodiment, the program includes information used for processing by an electronic computer that is equivalent to a program (such as data that is not a direct instruction to a computer but has properties that dictate computer processing).

[0045] Furthermore, in this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.

[0046] The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention. Furthermore, the processes described in the above embodiments may not only be executed in chronological order according to the order described, but may also be executed in parallel or individually depending on the processing capacity of the device that executes the processes or as needed.

Claims

1. A training data generation device comprising: a secondary training data set generation unit that defines a primary training data set as a set whose elements are primary training data consisting of a pair of primary input data, which is raw data to be input to a model, and primary labels, which are raw data indicating the evaluation results of the model's output for the primary input data; and a secondary training data set generation unit that generates, from the primary training data set, a secondary training data set whose elements are secondary input data, which is synthetic data to be input to a model generated from the primary input data extracted by a predetermined method; and a tertiary training data generation unit that generates, using the secondary training data set, a tertiary training data set whose elements are tertiary training data consisting of a pair of tertiary input data, which is synthetic data to be input to a model generated by a predetermined method, and tertiary labels, which are raw data indicating the evaluation results of the model's output for the tertiary input data; wherein the secondary training data set generation unit comprises: a probability distribution generation unit that generates a latent data space corresponding to the primary input data set from a primary input data set whose elements are primary input data constituting the primary training data, which are elements of the primary training data set, and generates a probability distribution in the latent data space; a gain function generation unit that generates a gain function used to extract primary input data that is expected to have a large utility in improving model performance, using the latent data space and the probability distribution; and a synthetic data generation unit that extracts primary input data for which synthetic data is to be generated based on the score of the gain function, and generates the secondary training data set whose elements are secondary input data generated from the extracted primary input data.

2. A training data generation device comprising: a primary training data set is a set whose elements are primary training data consisting of a set of primary input data, which is raw data to be input to a model, and primary labels, which are raw data indicating the evaluation results of the model's output for the primary input data; a secondary training data set generation unit that generates from the primary training data set a secondary training data set whose elements are secondary training data consisting of a set of secondary input data, which is synthetic data to be input to a model generated from primary input data extracted by a predetermined method and a set of primary labels corresponding to the primary input data, and secondary labels, which is synthetic data indicating the evaluation results of the model's output for the secondary input data; and a tertiary training data generation unit that generates a tertiary training data set whose elements are tertiary input data, which is synthetic data to be input to a model generated by a predetermined method using the secondary training data set, and tertiary labels, which are raw data indicating the evaluation results of the model's output for the tertiary input data, wherein the secondary training data set generation unit: a probability distribution generation unit that generates a latent data space corresponding to a primary input data set from a primary input data set whose elements are primary input data that constitute primary training data, which are elements of the primary training data set, and generates a probability distribution in the latent data space; a gain function generation unit that generates a gain function used to extract primary input data that is expected to have great utility in improving model performance, using the latent data space and the probability distribution; and a synthetic data generation unit that extracts primary input data for which synthetic data is to be generated based on a score of the gain function, and generates the secondary training data set whose elements are secondary training data generated from a set of the extracted primary input data and a primary label corresponding to the primary input data.

3. A training data generation device comprising: a primary training data set is a set whose elements are primary input data, which is raw data that serves as input to a model; a secondary training data set generation unit that generates, from the primary training data set, a secondary training data set whose elements are secondary input data, which is synthetic data that serves as input to a model generated from primary input data extracted by a predetermined method; and a tertiary training data generation unit that generates, using the secondary training data set, a tertiary training data set whose elements are tertiary training data consisting of a set of tertiary input data, which is synthetic data that serves as input to a model generated by a predetermined method, and tertiary labels, which are raw data that indicate evaluation results of the model output for the tertiary input data; wherein the secondary training data set generation unit comprises: a probability distribution generation unit that generates, from the primary training data set, a latent data space corresponding to the primary training data set and generates a probability distribution in the latent data space; and a gain function generation unit that generates a gain function using the latent data space and the probability distribution to extract primary input data that is expected to have a large utility for improving model performance. and a synthetic data generation unit that extracts primary input data for which synthetic data is to be generated based on the score of the gain function, and generates the secondary training data set whose elements are secondary input data generated from the extracted primary input data.

4. A training data generation method comprising: a secondary training data set generation step in which a primary training data set is defined as a set whose elements are primary input data, which is raw data to be input to a model; a training data generation device generating, from the primary training data set, a secondary training data set whose elements are secondary input data, which is synthetic data to be input to a model generated from primary input data extracted by a predetermined method; and a tertiary training data generation step in which the training data generation device generates, using the secondary training data set, a tertiary training data set whose elements are tertiary training data consisting of a set of tertiary input data, which is synthetic data to be input to a model generated by a predetermined method, and tertiary labels, which are raw data indicating the evaluation results of the model output for the tertiary input data, wherein the secondary training data set generation step comprises: a probability distribution generation step in which a latent data space corresponding to the primary training data set is generated from the primary training data set and a probability distribution in the latent data space is generated; and a gain function generation step in which a gain function is generated using the latent data space and the probability distribution to extract primary input data that is expected to be highly useful for improving model performance. and a synthetic data generation step of extracting primary input data for which synthetic data is to be generated based on the score of the gain function, and generating the secondary training data set whose elements are secondary input data generated from the extracted primary input data.

Images