Machine learning device, inference device and program

Abstract

To provide a learned model for inferring a state of a system.SOLUTION: A machine learning device 100 comprises: a first main component analysis unit 10 which generates two or more main components of learning data by performing main component analysis of the learning data; a first image data generation unit 20 which generates virtual image data by allocating the first main component of the two or more main components to X coordinates of an XY plane and allocating the second main component to Y coordinates of the XY plane; and a learning unit 30 which generates a learned model by performing machine learning with the image data as input data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a machine learning device, an inference device, and a trained model. [Background technology]

[0002] GoogLeNet, ResNet, Yolo, and other neural networks have been proposed for image recognition. These technologies use trained models based on deep learning to recognize specific objects in an image and infer what kind of object they are.

[0003] Principal Component Analysis (PCA) is also known as a method for finding components with large variance (hereinafter referred to as "principal components") that strongly represent the characteristics of the target data from correlated data with many dimensions. By using this method, it is possible to reduce the dimensions of data in machine learning.

[0004] Furthermore, a stochastic resonance phenomenon is known as a technique for extracting a hidden signal by applying noise to a weak signal buried in noise (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 6-174689 Summary of the Invention [Problem to be solved by the invention]

[0006] In general, building and verifying a neural network for the purpose of inferring the state of a system requires a great deal of man-hours, time, and skill. Here, we note that neural network technology for image recognition has been established, as mentioned above. If such neural networks for image recognition could be made to handle data other than images, it would be expected that building and verifying neural networks would become easier. To achieve this, the challenge is to convert non-image data into a format suitable for input to a neural network for image recognition.

[0007] The present invention has been made in light of these circumstances, and its purpose is to provide a technology for generating a trained model for inferring the state of a system. [Means for solving the problem]

[0008] In order to solve the above problem, a machine learning device according to one embodiment of the present invention includes: a first principal component analysis unit that generates two or more principal components of the training data by performing principal component analysis on the training data; a first image data generation unit that generates virtual image data by assigning a first principal component to an X coordinate on an XY plane and a second principal component to a Y coordinate on the XY plane, among the two or more generated principal components; and a learning unit that generates a trained model by performing learning using a neural network with the image data as input data.

[0009] Another aspect of the present invention is an inference device that includes a second principal component analysis unit that generates two or more principal components of the observation data by performing principal component analysis on the observation data, a second image data generation unit that generates virtual image data by assigning a first principal component of the two or more principal components to the X coordinate on an XY plane and a second principal component to the Y coordinate on the XY plane, and an inference unit that performs inference using a trained model with the image data as input data.

[0010] Yet another aspect of the present invention is a trained model, which is generated by: a principal component analysis step of generating two or more principal components of the training data by performing principal component analysis on the training data; an image data generation step of generating virtual image data by assigning a first principal component of the generated two or more principal components to the X coordinate on an XY plane and a second principal component to the Y coordinate on the XY plane; and machine learning using the image data as input data.

[0011] Any combination of the above components, and any transformation of the present invention into a method, device, system, recording medium, computer program, etc., are also valid aspects of the present invention. [Effects of the Invention]

[0012] According to the present invention, a trained model for inferring the state of a system can be generated. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a functional block diagram of a machine learning device according to a first embodiment. [Figure 2] 2 is a diagram illustrating an example of a set of data input to a first principal component analysis unit of the machine learning device of FIG. 1. FIG. [Figure 3] These are virtual images drawn using image data generated by the first image data generation unit of the machine learning device in Figure 1. (a) is a virtual image in which the source of the odor is flower A, (b) is a virtual image in which the source of the odor is perfume B, and (c) is a virtual image in which the source of the odor is fabric softener C. [Figure 4] FIG. 2 is a schematic diagram of the first principal component, the second principal component, the third principal component, the fourth principal component, and the fifth principal component assigned to the XY plane by the first image data generating unit of the machine learning device of FIG. 1. [Figure 5] FIG. 10 is a functional block diagram of a machine learning device according to a fourth embodiment. [Figure 6]These are portions of virtual images drawn using image data generated by the first image data generation unit of the machine learning device in Fig. 5. (a) is a virtual image in which the source of the odor is flower A, (b) is a virtual image in which the source of the odor is perfume B, and (c) is a virtual image in which the source of the odor is fabric softener C. [Figure 7] FIG. 10 is a functional block diagram of an inference device according to a fifth embodiment. [Figure 8] FIG. 13 is a functional block diagram of an inference system according to a sixth embodiment. [Figure 9] 13 is a flowchart illustrating a procedure for generating a trained model according to the seventh embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] The present invention will be described below based on preferred embodiments with reference to the drawings. In the embodiments and modifications, identical or equivalent components and parts are designated by the same reference numerals, and redundant description will be omitted where appropriate. The dimensions of the components in the drawings are enlarged or reduced as appropriate for ease of understanding. Some elements that are not important for explaining the embodiments are omitted from the drawings. Terms including ordinal numbers such as "first" and "second" are used to describe various components, but these terms are used only to distinguish one component from another and do not limit the components.

[0015] [First embodiment] 1 is a functional block diagram of a machine learning device 100 according to a first embodiment. The machine learning device 100 includes a first principal component analysis unit 10, a first image data generation unit 20, and a learning unit 30.

[0016] External data is input to the first principal component analysis unit 10 as training data. The first principal component analysis unit 10 performs principal component analysis on the data to generate two or more principal components of the training data. The first image data generation unit 20 generates virtual image data by assigning the first principal component of the two or more principal components generated by the first principal component analysis unit 10 to the X coordinate on the XY plane and the second principal component to the Y coordinate on the XY plane. The learning unit 30 generates a trained model by performing training using a neural network with the image data generated by the first image data generation unit 20 as input data.

[0017] Below, we will use odor data as an example of training data to specifically explain the operation of the machine learning device 100, which generates a trained model for inferring the source of an odor present in space and the state of a system.

[0018] Smell data from eight smell sensors (sensors 1 to 8) arranged outside the machine learning device 100 is input to the first principal component analysis unit 10. Here, each smell sensor is assumed to have high sensitivity to the following smell components. Sensor 1: Methane Sensor 2: Ethane Sensor 3: Propane Sensor 4: Acetic acid Sensor 5: Ethanol Sensor 6: Toluene Sensor 7: Hydrogen sulfide Sensor 8: Acetaldehyde In other words, the intensities of the eight types of odor components mentioned above can be considered to be input to the first principal component analysis unit 10. The first principal component analysis unit 10 performs principal component analysis on these data to generate first and second principal components of the odor data according to the odor source (flower A, perfume B, fabric softener C, mouthwash D, etc.). As is well known, the first principal component is generated as a linear combination of odor components that maximizes the variance of the data for each odor source. The second principal component is generated as a linear combination of odor components that maximizes the variance of the data projected onto an axis perpendicular to the first principal component axis.

[0019] FIG. 2 shows sets of odor data input from sensors 1 to 8 to first principal component analysis unit 10. This figure shows the case where the odor source is flower A. Each row in FIG. 2 shows a set of data acquired by each sensor in one observation. In this example, each sensor acquired a total of 2000 pieces of training data from the first to the 2000th observation. Each column in FIG. 2 shows a set of data acquired by each sensor in a total of 2000 observations. However, FIG. 2 only shows the training data from the first observation, and omits data values ​​from the second observation onwards.

[0020] The first principal component analysis unit 10 extracts appropriate subsets of data from the data set in Fig. 2, and generates a first principal component and a second principal component for each subset. The subsets may be extracted, for example, as follows. (Example 1) If a subset is extracted as shown below, 200 pairs of the first and second principal components generated from 10 sets of training data will be generated. Subset 1: 1st training data to 10th training data Subset 2: 11th training data to 20th training data … Subset 200: 1991st training data to 2000th training data (Example 2) When subsets are extracted as shown below, 40 pairs of the first and second principal components are generated from 50 sets of training data. Subset 1: 1st training data to 50th training data Subset 2: 51st training data to 100th training data … Subset 40: 1951st training data to 2000th training data The method for extracting the subset is arbitrary and may be determined based on the amount of calculation and accuracy.

[0021] The first image data generation unit 20 assigns the first principal component of the two or more principal components generated by the first principal component analysis unit 10 to the X coordinate of the XY plane, and the second principal component to the Y coordinate of the XY plane. Next, the first image data generation unit 20 draws points on the XY plane for each extracted subset, with the first principal component as the X coordinate and the second principal component as the Y coordinate. Here, the pixel value of each point is an arbitrary value indicating the presence or absence of a point, such as a binary value. The first image data generation unit 20 generates a single piece of image data by drawing the first and second principal components of multiple subsets on the XY plane. This makes it possible to draw a "virtual image" corresponding to the source of the odor on the XY plane. In other words, the first image data generation unit 20 generates virtual image data corresponding to the source of the odor.

[0022] 3 shows virtual images drawn using image data generated by the first image data generation unit 20. (a) is a virtual image in which the source of the odor is flower A, (b) is a virtual image in which the source of the odor is perfume B, and (c) is a virtual image in which the source of the odor is fabric softener C. These images were generated by plotting multiple points on the XY plane through principal component analysis of multiple subsets.

[0023] In this way, the odor data is image data corresponding to the source of the odor, making it suitable as input data for a neural network for image recognition.

[0024] The learning unit 30 performs learning using a neural network with the image data generated by the first image data generation unit 20 as input data. The learning may be supervised learning using training data in which the input data is labeled with the source of the odor, or may be unsupervised learning. In this way, when odor data is given, the first image data generation unit 20 generates a trained model for inferring the source of the odor and the state of the system. The neural network used by the learning unit 30 may be any suitable neural network for image recognition, such as GoogLeNet, ResNet, or Yolo.

[0025] The above describes an example in which the learning unit 30 performs learning using a neural network. However, the present invention is not limited to this, and the learning unit 30 may generate a trained model using a machine learning method other than a neural network.

[0026] In this way, the original odor data is converted into image data, making the data suitable for a neural network for image recognition. That is, according to this embodiment, a trained model for inferring the state of a system from arbitrary training data can be generated using a neural network suitable for image processing.

[0027] [Second embodiment] In one embodiment, the first principal component analysis unit 10 of the machine learning device 100 in Fig. 1 generates three or more principal components of the training data by performing principal component analysis on the training data. The third principal component among the generated three or more principal components is generated as a linear combination of odor components that maximizes the variance of the data projected onto an axis orthogonal to the first and second principal component axes.

[0028] The first image data generation unit 20 assigns the first principal component of the three or more principal components generated by the first principal component analysis unit 10 to the X coordinate of the XY plane, the second principal component to the Y coordinate of the XY plane, and in addition, assigns the third principal component to the pixel value at each point on the XY plane. This makes it possible to draw a virtual image (i.e., grayscale) on the XY plane, including shading information corresponding to the odor source. That is, the first image data generation unit 20 generates virtual image data including shading information corresponding to the odor source.

[0029] Before generating virtual image data, the first image data generating unit 20 may normalize the third principal component according to the resolution of the image. For example, the first image data generating unit 20 may normalize the data range of the third principal component to 0-255, 0-511, or 0-1023.

[0030] The learning unit 30 performs learning using a neural network with the grayscale image data generated by the first image data generating unit 20 as input data.

[0031] According to this embodiment, since the image data input to the neural network for image recognition contains shading information, it is possible to generate a trained model for inferring the state of the system with higher accuracy.

[0032] [Third embodiment] In one embodiment, the first principal component analysis unit 10 of the machine learning device 100 in Fig. 1 generates four or more principal components of the training data by performing principal component analysis on the training data. The fourth principal component of the generated four or more principal components is generated as a linear combination of odor components that maximizes the variance of data projected onto an axis orthogonal to the first, second, and third principal component axes. Similarly, the fifth principal component is generated as a linear combination of odor components that maximizes the variance of data projected onto an axis orthogonal to the first, second, third, and fourth principal component axes.

[0033] The first image data generation unit 20 assigns the first principal component of the four or more principal components generated by the first principal component analysis unit 10 to the X coordinate of the XY plane, assigns the second principal component to the Y coordinate of the XY plane, and in addition, assigns each of the third and subsequent principal components to the pixel value at each point on the XY plane and each color component value in the color space at each point on the XY plane.

[0034] For example, the first image data generation unit 20 may assign the third principal component to the R value of the RGB color space at each point on the XY plane, the fourth principal component to the G value of the RGB color space at each point on the XY plane, and the fifth principal component to the B value of the RGB color space at each point on the XY plane. This makes it possible to draw a virtual image on the XY plane that includes color information corresponding to the odor source (i.e., a full-color image). That is, the first image data generation unit 20 generates virtual image data that includes color information corresponding to the odor source.

[0035] In the above description, the RGB color space has been used as an example of the color space to which the third and subsequent principal components are assigned, but the color space is not limited to this and may be any color space such as CMYK, HSL, etc. Furthermore, when the first principal component analysis unit 10 generates principal components up to the Nth principal component, the (N-2)th principal component and subsequent principal components may be assigned to each color component in the color space.

[0036] Before generating virtual image data, the first image data generating unit 20 may normalize each of the third and subsequent principal components according to the color resolution of the image. For example, the first image data generating unit 20 may normalize the data ranges of the third, fourth, and fifth principal components to 0-255, 0-511, or 0-1023.

[0037] FIG. 4 schematically shows the first, second, third, fourth and fifth principal components assigned to the XY plane by the first image data generating section 20. As shown in FIG.

[0038] The learning unit 30 performs learning using a neural network with the color image data generated by the first image data generating unit 20 as input data.

[0039] According to this embodiment, since the image data input to the neural network for image recognition contains color information, it is possible to generate a trained model that can infer the state of the system with even greater accuracy.

[0040] [Fourth embodiment] 5 is a functional block diagram of a machine learning device 110 according to the fourth embodiment. The machine learning device 110 includes a first principal component analysis unit 10, a first image data generation unit 20, a learning unit 30, and a noise application unit 40. That is, the machine learning device 110 includes the noise application unit 40 in addition to the configuration of the machine learning device 100 in FIG.

[0041] The noise applying section 40 applies noise to points around the point on the XY plane for which the virtual image data has been generated by the first image data generating section 20.

[0042] The virtual image corresponding to the odor source drawn using the above method may be considered to be noise if, for example, the number of points corresponding to pixels is not sufficient. In this case, adding noise around the points corresponding to the pixels causes the stochastic resonance phenomenon, improving the signal-to-noise ratio.

[0043] 6 shows a portion of a virtual image drawn using image data generated by the first image data generation unit 20 of the machine learning device 110. (a) is a virtual image in which the source of the odor is flower A, (b) is a virtual image in which the source of the odor is perfume B, and (c) is a virtual image in which the source of the odor is fabric softener C. The noise application unit 40 adds noise to points (areas surrounded by circles) surrounding points drawn on the XY plane by the first image data generation unit 20.

[0044] According to this embodiment, by applying noise to points surrounding each point that constitutes a virtual image drawn on the XY plane, the accuracy of the trained model can be further improved by the stochastic resonance phenomenon.

[0045] The above describes an example of a trained model for inferring the source of an odor and the state of a system based on odor data. However, embodiments may generate trained models from data other than this. For example, embodiments may generate trained models for inferring or predicting weather conditions and air pollution based on meteorological data such as temperature, humidity, atmospheric pressure, atmospheric carbon dioxide concentration, and atmospheric fine particulate matter concentration.

[0046] The various processes of the machine learning device described above can be realized not only as a device using hardware such as a CPU and memory, but also as firmware stored in a ROM (read-only memory) or flash memory, or software for a computer, etc. The firmware program or software program can be provided by being recorded on a computer-readable recording medium, or can be transmitted and received with a server via a wired or wireless network, or can be transmitted and received as data broadcasting via terrestrial or satellite digital broadcasting.

[0047] [Fifth embodiment] 7 is a functional block diagram of an inference device 200 according to the fifth embodiment. The inference device 200 includes a second principal component analysis unit 12, a second image data generation unit 22, and an inference unit 50.

[0048] External observation data is input to the second principal component analysis unit 12. The second principal component analysis unit 12 performs principal component analysis on the observation data to generate two or more principal components of the observation data. The second image data generation unit 22 generates virtual image data for inference by assigning a first principal component of the two or more principal components generated by the second principal component analysis unit 12 to the X coordinate on the XY plane and assigning the second principal component to the Y coordinate on the XY plane. The inference unit 50 estimates the state of the system by applying a trained model to the image data generated by the second image data generation unit 22.

[0049] The trained model may be generated by, for example, the above-described machine learning device 100 or machine learning device 110. The trained model may be stored in a storage unit provided inside the inference device 200, or may be downloaded from a database or the like external to the inference device 200.

[0050] According to this embodiment, the state of a system can be estimated from observation data using a trained model.

[0051] [Sixth embodiment] FIG. 8 is a functional block diagram of an inference system 300 according to the sixth embodiment. The inference system 300 includes a machine learning device 100 and an inference device 200. The machine learning device 100 is the same as that described in the first embodiment, and includes a first principal component analysis unit 10, a first image data generation unit 20, and a learning unit 30. The operation of the machine learning device 100 is as described in the first embodiment. The inference device 200 is the same as that described in the fifth embodiment, and includes a second principal component analysis unit 12, a second image data generation unit 22, and an inference unit 50. The operation of the inference device 200 is as described in the fifth embodiment.

[0052] The first principal component analysis unit and the second principal component analysis unit, and the first image data generation unit and the second image data generation unit may be realized by the same or different hardware, or may be implemented as the same or different software. That is, they may be configured by a common principal component analysis unit and image data generation unit.

[0053] According to this embodiment, it is possible to estimate the state of a system from observation data using a trained model generated by a neural network suitable for image processing.

[0054] [Seventh embodiment] FIG. 9 is a flowchart illustrating a procedure for generating a trained model according to the seventh embodiment.

[0055] First, in this method, a first principal component analysis unit is used to perform principal component analysis on the training data to generate two or more principal components of the training data (S1).

[0056] Next, the method uses a first image data generation unit to generate virtual image data by assigning the first principal component of the two or more principal components generated in S1 to the X coordinate on the XY plane and the second principal component to the Y coordinate on the XY plane (S2).

[0057] Next, this method uses a learning unit to perform learning using a neural network with the image data generated in S2 as input data, thereby generating a trained model (S3).

[0058] According to this embodiment, a trained model for inferring the state of a system can be generated using a neural network suitable for image processing.

[0059] [Eighth embodiment] The eighth embodiment is a computer program that causes a computer to execute the following steps: generating two or more principal components of the training data by performing principal component analysis on the training data; generating virtual image data by assigning a first principal component of the generated two or more principal components to the X coordinate on an XY plane and assigning a second principal component to the Y coordinate on the XY plane; and performing training using a neural network with the generated image data as input data to generate a trained model.

[0060] According to this embodiment, a program that generates a trained model for inferring the state of a system using a neural network suitable for image processing can be implemented as software.

[0061] The present invention has been described above based on the embodiments. The embodiments are merely examples, and it will be understood by those skilled in the art that various modifications are possible in the combination of the components and treatment processes, and that such modifications are also within the scope of the present invention. [Explanation of symbols]

[0062] 10...First principal component analysis unit, 12...Second principal component analysis unit, 20...First image data generation unit, 22...Second image data generation unit, 30...Learning unit, 40...Noise application unit, 50...Inference unit, S1...Step of performing principal component analysis on training data, S2...Step of generating virtual image data, S3...Step of performing learning to generate a trained model, 100...Machine learning device, 110...Machine learning device, 200...Inference device, 300...Inference system.

Claims

1. a first principal component analysis unit that generates two or more principal components of the training data by performing principal component analysis on the training data; a first image data generating unit that generates virtual image data by allocating a first principal component of the two or more principal components to an X coordinate on an XY plane and a second principal component to a Y coordinate on the XY plane; A machine learning device comprising: a learning unit that generates a trained model by performing machine learning using the image data as input data.

2. the first principal component analysis unit generates three or more principal components of the training data by performing principal component analysis on the training data; the first image data generation unit assigns each of the three or more principal components, starting from a third principal component, to a pixel value at each point on an XY plane; The machine learning device according to claim 1 , wherein the learning unit generates a trained model by performing machine learning using the image data as input data.

3. the first principal component analysis unit generates four or more principal components of the training data by performing principal component analysis on the training data; the first image data generation unit assigns each of the third and subsequent principal components among the four or more principal components to a pixel value at each point on an XY plane and to each color component value of a color space at each point on the XY plane; The machine learning device according to claim 1 , wherein the learning unit generates a trained model by performing machine learning using the image data as input data.

4. 4. The machine learning device according to claim 1, further comprising a noise applying unit that applies noise to points around a point on an XY plane where the virtual image data is generated.

5. a second principal component analysis unit that generates two or more principal components of the observation data by performing principal component analysis on the observation data; a second image data generating unit that generates virtual image data by allocating a first principal component of the two or more principal components to an X coordinate on an XY plane and a second principal component to a Y coordinate on the XY plane; an inference unit that performs inference using a trained model with the image data as input data.

6. A program executed by a computer of a machine learning device, a principal component analysis step of performing principal component analysis on the training data to generate two or more principal components of the training data; an image data generating step of generating virtual image data by allocating a first principal component of the two or more principal components to an X coordinate on an XY plane and allocating a second principal component to a Y coordinate on the XY plane; and generating a trained model by performing machine learning using the image data as input data.

Images