Measurement device, communication system, measurement method, and program

Abstract

The purpose of the present disclosure is to more accurately measure changes in the behavior of a machine learning model without having to examine the output of the machine learning model for every input thereto.　For this purpose, the present disclosure regards machine learning models as functions, defines, as a similarity, the sum of the differences between the output data of two machine learning models for the same input data thereto, and measures this similarity using three measurement methods (input range restriction, similarity estimation, function approximation) instead of examining the output for each input. This has the advantage that it is possible to measure changes in the behavior of a prescribed machine learning model without having to examine the output for every input, and that an appropriate one of the above three measurement methods may be used to more accurately measure changes in the behavior of the machine learning model.

Description

[Technical Field]

[0001] The present disclosure relates to techniques for measuring changes in the behavior of machine learning models. [Background technology]

[0002] Among the many machine learning models (AI: Artificial Intelligence), some are constantly changing due to continuous machine learning. This constant change is a characteristic of machine learning models that is not found in conventional software, and machine learning models may be changing without you even realizing it. Known methods for understanding changes in the behavior (state) of machine learning models include visualizing the internal parameters of the machine learning model (such as weights within the model) and visualizing the inference accuracy of the machine learning model by inputting test data into the machine learning model (see Non-Patent Document 1). [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Hohman, Fred, et al. "Visual analytics in deep learning: An interrogative survey for the next frontiers." IEEE transactions on visualization and computer graphics 25.8 (2018): 2674-2693. Summary of the Invention [Problem to be solved by the invention]

[0004] However, simply visualizing the internal parameters of a machine learning model does not allow us to understand how the model's behavior affects its output. Furthermore, even if we visualize the inference accuracy of a machine learning model, this is insufficient to understand the model's behavior, since even machine learning models with similar inference accuracy can differ in the targets they can infer and the way they make mistakes. Furthermore, investigating the output for every input in order to understand the state of a machine learning model would require a huge amount of work, making it unrealistic.

[0005] The present invention has been made in consideration of the above points, and aims to measure changes in the behavior of a machine learning model more accurately without examining the outputs for all inputs to the machine learning model. [Means for solving the problem]

[0006] In order to solve the above problem, the invention of claim 1 is a measurement device for measuring changes in behavior of a machine learning model, comprising: an input unit capable of inputting data in a first input range and data in a second input range; an input range determination unit which, when the data in the first input range and data in the second input range are input by the input unit, determines whether the data in the first input range and the data in the second input range are identical, and, when determined to be identical, sets the data in the first input range and the data in the second input range to data in the same predetermined input range; a set setting unit which sets the data in the predetermined input range itself as data in an input set, or sets a value sampled from the data in the predetermined input range as data in the input set; and a function setting unit which inputs data of each element of the set to a first function related to a first machine learning model and a second function related to a second machine learning model. a set element input unit that inputs data of each element of the set to each of a first approximation function that approximates a decision tree or rule simpler than the first machine learning model and a second approximation function that approximates a decision tree or rule simpler than the second machine learning model; a similarity measurement unit that, when the data of each element of the set is input to each of the first function and the second function by the set element input unit, measures a similarity between first output data from the first function and second output data from the second function; or, when the data of each element of the set is input to each of the first approximation function and the second approximation function by the set element input unit, measures a similarity between first output data from the first approximation function and second output data from the second approximation function; and an output unit that outputs a result related to the similarity measured by the similarity measurement unit. [Effects of the Invention]

[0007] As described above, the present invention has the effect of being able to measure changes in the behavior of a machine learning model more accurately without examining the outputs for all inputs to the machine learning model. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a diagram illustrating the overall configuration of a communication system according to an embodiment of the present invention. [Figure 2] 1 is a diagram showing the electrical hardware configuration of a measurement device and a communication terminal according to this embodiment. [Figure 3] FIG. 2 is a functional configuration diagram of the measurement device. [Figure 4] 4 is a flowchart showing a process or operation executed by the measurement device. [Figure 5] 4 is a flowchart showing a process or operation executed by the measurement device. [Figure 6] 4 is a flowchart showing a process or operation executed by the measurement device. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0010] [System configuration of the embodiment] First, the overall configuration of the communication system of this embodiment will be explained with reference to Fig. 1. Fig. 1 is a diagram showing the overall configuration of the communication system according to this embodiment.

[0011] As shown in Fig. 1, a communication system 1 of this embodiment is constructed by a measurement device 3 and a communication terminal 5. The communication terminal 5 is managed and used by a user. The user refers to the output result of the measurement device and determines the subsequent action to be taken.

[0012] Furthermore, the measurement device 3 and the communication terminal 5 can communicate with each other via a communication network 100 such as the Internet. The connection form of the communication network 100 may be either wireless or wired.

[0013] The measurement device 3 is configured by one or more computers. When the measurement device 3 is configured by multiple computers, it may be referred to as a "measurement device" or a "measurement system."

[0014] The measurement device 3 regards the input and output of the machine learning model as a function, defines the similarity, and then measures an alternative similarity so that it is not necessary to check the output for every input.

[0015] The communication terminal 5 is a computer, and a notebook computer is shown as an example in Fig. 1. In Fig. 1, a user operates the communication terminal 5. Note that processing may be performed by the measurement device 3 alone, without using the communication terminal 5.

[0016] [Hardware configuration] <Hardware configuration of the measurement device> Next, the electrical hardware configuration of the measurement device 3 will be described with reference to Fig. 2. Fig. 2 is a diagram showing the electrical hardware configuration of the measurement device.

[0017] As shown in FIG. 2, the measuring device 3 is a computer and includes a CPU (Central Processing Unit) 301, a ROM (Read Only Memory) 302, a RAM (Random Access Memory) 303, an SSD (Solid State Drive) 304, an external device connection I / F (Interface) 305, a network I / F 306, a media I / F 309, and a bus line 310.

[0018] Of these, the CPU 301 controls the overall operation of the measuring device 3. The ROM 302 stores programs such as an IPL (Initial Program Loader) used to drive the CPU 301. The RAM 303 is used as a work area for the CPU 301.

[0019] The SSD 304 reads or writes various data under the control of the CPU 301. Note that instead of the SSD 304, a hard disk drive (HDD) may be used.

[0020] The external device connection I / F 305 is an interface for connecting various external devices, such as a display, a speaker, a keyboard, a mouse, a USB (Universal Serial Bus) memory, and a printer.

[0021] The network I / F 306 is an interface for performing data communication via the communication network 100 .

[0022] The media I / F 309 controls reading and writing (storing) of data from and to a recording medium 309m such as a flash memory, etc. The recording medium 309m includes a DVD (Digital Versatile Disc) and a Blu-ray Disc (registered trademark).

[0023] The bus line 310 is an address bus, a data bus, etc. for electrically connecting the components such as the CPU 301 shown in FIG.

[0024] The electrical hardware configuration of the communication terminal 5 is basically the same as that of the measurement device 3, and therefore a description thereof will be omitted.

[0025] [Measurement development technology details] Next, the details of the developed technology for measurement using the measurement device will be explained.

[0026] <Definition of similarity> The input and output of the machine learning model (machine learning model) are considered as functions, and the two machine learning models whose similarity is measured are called f1 and f2. The input (data) and output (data) of the machine learning model are multivariate vectors.

[0027]

number

[0028]

number

[0029]

number

[0030]

number

[0031]

number

[0032]

number

[0033] <Measurement of similarity> However, since it is unrealistic to grasp all pairs of input and output (data) of a machine learning model, it is difficult to directly measure the similarity described above.

[0034] Therefore, in this embodiment, the following three measurement methods (measurement techniques) are used to obtain similarity that replaces the above equation (Equation 2). The three measurement techniques may be used alone or in combination of two or more. Input range limited ·Similarity estimation Function approximation <Input range limit> The range of similarity calculation is limited to set D, and the similarity in set D is measured. That is, the similarity is measured by substituting the following formula.

[0035]

number

[0036]

number

[0037] Since the actual inputs are discrete values, the integral value may be replaced by a simple sum. If the input ranges of the two machine learning models are different, a domain conversion process may be performed to convert domain D1 to D2 (or domain D2 to D1) to align the domains. This conversion process corresponds to, for example, converting the size of an image.

[0038] <Similarity estimation> The desired similarity is estimated from a set of sampled input and output data using statistical methods. Randomly extract n differences E(f1(x), f2(x)) for a certain input data x, and calculate the average.

[0039]

number

[0040] Assuming that the population of differences follows a normal distribution, the similarity can be estimated using the following formula:

[0041]

number

[0042]

number

[0043] As shown in FIG. 3, the measurement device 3 includes an input unit 31, a measurement method determination unit 32, an input range acquisition unit 33, an input range determination unit 34, a domain conversion unit 35, a set setting unit 36, a set element input unit 37, a similarity measurement unit 38, and an output unit 39. Each of these units has a function realized by an instruction from the CPU 301 in FIG. 2 based on a program. Furthermore, machine learning models 41 and 42 are stored in the RAM 303 or the SSD 304. Note that at least one of the machine learning models 41 and 42 may be stored not in the measurement device 3 but in a device independent of the measurement device 3, and the measurement device 3 may use at least one of the machine learning models 41 and 42 via the communication network 100.

[0044] The input unit 31 inputs input content (designation of measurement method) received from the communication terminal 5 or input content (designation of measurement method) received directly by the input unit 31 via its own device (measurement device 3). The designated measurement method is at least one of three types of measurement methods (input range limitation, similarity estimation, or function approximation). If "input range limitation" is designated, the input content includes data D1 of a first input range and data D2 of a second input range related to the comparison target. If "function approximation" is designated, the input content includes approximate functions f'1 and f'2. Note that the data in the input range indicates data generated or acquired within a predetermined range (a predetermined period, a predetermined classification, etc.). The data in the input range may not be the data itself generated or acquired within the predetermined range, but may be data indicating the range conditions. For example, the range conditions may be such that x is between 1 and 500.

[0045] The measurement method determination unit 32 determines which of the three measurement methods has been designated based on the input contents.

[0046] If the input unit 31 does not input data D1 of the first input range and data D2 of the second input range because "input range limitation" is not specified, the input range acquisition unit 33 acquires data d1 of the predetermined first input range from the machine learning model 41 and acquires data d2 of the predetermined second input range from the machine learning model 42.

[0047] When "input range limitation" is specified and data D1 of the first input range and data D2 of the second input range are input by the input unit 31, the input range determination unit 34 determines whether the data D1 of the first input range and data D2 of the second input range are the same, and if it is determined that they are the same, it sets the data D1 of the first input range and data D2 of the second input range to the same predetermined input range data D. Furthermore, when "input range limitation" is not specified and data d1 of the predetermined first input range and data d2 of the predetermined second input range acquired by the input range acquisition unit 33 are the same, it sets the data d1 of the predetermined first input range and data d2 of the predetermined second input range to the same predetermined input range data D.

[0048] When the first input range data D1 and the second input range data D2 are input by the input unit 31 and the first input range data D1 and the second input range data D2 are not identical, the domain conversion unit 35 converts the domain of the first input range data D1 or the second input range data D2 to align the first input range and the second input range to form predetermined input range data D. Furthermore, when the predetermined first input range data d1 and the predetermined second input range data d2 are not identical, the domain conversion unit 35 converts the domain of the predetermined first input range data d1 or the predetermined second input range data d2 to align the predetermined first input range and the predetermined second input range to form predetermined input range data D.

[0049] If "similarity estimation" is specified, the set setting unit 36 ​​sets the value x sampled from the data D in a predetermined input range as the data X of the input set. If "similarity estimation" is not specified, the set setting unit 36 ​​sets the data D itself in the predetermined input range as the data X of the input set.

[0050] If "function approximation" is not specified, the set element input unit 37 inputs data x of each element of the set to each of a first function related to the first machine learning model and a second function related to the second machine learning model. If "function approximation" is specified, the set element input unit 37 inputs data x of each element of the set to each of a first approximation function that approximates a decision tree or rule simpler than machine learning model 41 and a second approximation function that approximates a decision tree or rule simpler than machine learning model 42.

[0051] When set element input unit 37 inputs data x of each element of the set to each of the first function and the second function, similarity measurement unit 38 measures the similarity between the first output data from the first function and the second output data from the second function. Furthermore, when set element input unit 37 inputs data x of each element of the set to each of the first approximation function and the second approximation function, similarity measurement unit 38 measures the similarity between the first output data from the first approximation function and the second output data from the second approximation function.

[0052] The output unit 39 outputs information on the measurement result indicating the similarity measured by the similarity measurement unit 38 from the measurement device 3. Examples of output include displaying the information on a display connected to the external device connection I / F 305 in Fig. 2, or transmitting the information to an external device such as the communication terminal 5 via the network I / F 306.

[0053] [Processing or operation of the measuring device] Next, the processing or operation of the measurement device 3 will be described with reference to Figures 4 to 6. Figures 4 to 6 are flowcharts showing the processing or operation executed by the measurement device.

[0054] First, based on the input content received by the input unit 31 from the communication terminal 5 or the input content received directly by the input unit 31 for its own device (measurement device 3), the measurement method determination unit 32 determines which of the three types of similarity measurement methods (input range limitation, similarity estimation, or function approximation) has been designated (or selected) (hereinafter, this process will be referred to as "measurement method determination process"). In this case, at least one of the three types of measurement methods is accepted.

[0055] S12: In the measurement method determination process, if the measurement method determination unit 32 does not determine that "input range limitation" has been specified (S11; NO), the input range acquisition unit 33 acquires input range data d1 from the machine learning model 41 and acquires input range data d2 from the machine learning model 42. Then, the process proceeds to S13.

[0056] S13: If the measurement method determination unit 32 determines that the designation of "input range limitation" has been received (S11; YES), the input range determination unit 34 determines whether the input range data D1, D2 input by the input unit 31 are the same. Alternatively, if the measurement method determination unit 32 does not determine that the designation of "input range limitation" has been received (S11; NO), the input range determination unit 34 determines whether the input range data d1, d2 input by the input range acquisition unit 33 in step S12 are the same.

[0057] S14: If the input range data D1 and D2 are the same, or the input range data d1 and d2 are the same (S13; YES), the input range determination unit 34 determines the input range data D1 and D2 or the input range data d1 and d2 as the specified input range data D.

[0058] S15: When the input range data D1 and D2 are not the same (different), or when the input range data d1 and d2 are not the same (different) (S13; YES), the domain conversion unit 35 converts the domain of either one of the input range data D1 and D2 or one of the input range data d1 and d2 to align the input range data D1 and D2, or align the input range data d1 and d2, and set this as the predetermined input range data D. Note that the domain conversion unit 35 may convert each domain of the input range data D1 and D2 to align them to the input range data D, or may convert each domain of the input range data d1 and d2 to align them to the input range data D.

[0059] S22: Furthermore, in the measurement method determination process, as shown in FIG. 5, if the measurement method determination unit 32 determines that the designation of “similarity measurement” has been received (S21; YES), the set setting unit 36 ​​sets the value x sampled from the input range data D aligned by the input range determination unit 34 as the input set X, or sets the value x sampled from the input range data D aligned by the domain conversion unit 35 as the input set X.

[0060] S23: On the other hand, if the measurement method determination unit 32 does not determine that a "similarity measurement" designation has been received (if it is determined that it has not been received) in the measurement method determination process (S21; NO), the set setting unit 36 ​​sets the data D of the input range aligned by the input range determination unit 34 or the data D of the input range aligned by the domain conversion unit 35 as the data X of the input set.

[0061] S32: Furthermore, in the measurement method determination process, as shown in FIG. 6, if the measurement method determination unit 32 determines that "function approximation" has been specified (S31; YES), the set element input unit 37 inputs the data x of each element of the set data X into the approximation function [f'1(x), f'2(x)] to obtain each output [y'1, y'2].

[0062] S33: On the other hand, in the measurement method determination process, as shown in FIG. 6, if the measurement method determination unit 32 does not determine that the designation of "function approximation" has been received (determines that it has not been received) (S33; NO), the set element input unit 37 inputs the data x of each element of the data X of the set into the functions [f1(x), f2(x)] of the machine learning models 41 and 42 to obtain each output [y1, y2].

[0063] S34: The similarity measurement unit 38 measures the similarity between each of the output data [y1, y2], or measures the similarity between each of the output data [y1, y2].

[0064] S35: The output unit 39 outputs the result of the similarity measured by the similarity measurement unit .

[0065] This completes the description of the processing or operation of the measuring device.

[0066] [Effects of the embodiment] As described above, according to this embodiment, a machine learning model is regarded as a function, the sum of the differences in output data for the same input data to two machine learning models is defined as similarity, and instead of examining the output for the input, this similarity is measured using three measurement techniques (input range limitation, similarity estimation, and function approximation). This makes it possible to measure changes in the behavior of a given machine learning model without examining the outputs for all inputs, and furthermore, by appropriately using any of the three measurement techniques, it is possible to more accurately measure changes in the behavior of the machine learning model.

[0067] 〔supplement〕 The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations). (1) The measuring device 3 can be realized by a computer and a program, but this program can also be recorded on a (non-transitory) recording medium or provided via the communication network 100. (2) In the above embodiment, a notebook computer is shown as an example of the communication terminal 5, but this is not limited to this and may be, for example, a desktop computer, a tablet terminal, a smartphone, a smartwatch, a car navigation device, a refrigerator, a microwave oven, etc. (3) The CPU 301 may be multiple, not just single. [Explanation of symbols]

[0068] 1. Communication Systems 3. Measuring equipment 5. Communication terminals 31 Input section 32 Measurement method judgment section 33 Input range acquisition section 34 Input range determination section 35 Domain conversion section 36 Set Setting Section 37 Set element input section 38 Similarity measurement part 39 Output section 41 Machine Learning Model (Example of the first machine learning model) 42 Machine Learning Model (An example of a second machine learning model)

Claims

1. A measurement device for measuring changes in the behavior of a machine learning model, an input unit capable of inputting data in a first input range and data in a second input range; an input range determination unit that, when the first input range data and the second input range data are input by the input unit, determines whether the first input range data and the second input range data are identical, and, when it is determined that they are identical, sets the first input range data and the second input range data as data of the same predetermined input range; a set setting unit that sets the data in the predetermined input range itself as data in the set for input, or sets values ​​sampled from the data in the predetermined input range as data in the set for input; a set element input unit that inputs data of each element of the set to each of a first function related to a first machine learning model and a second function related to a second machine learning model, or that inputs data of each element of the set to each of a first approximation function that approximates a decision tree or rule simpler than the first machine learning model and a second approximation function that approximates a decision tree or rule simpler than the second machine learning model; a similarity measurement unit that measures a similarity between first output data from the first function and second output data from the second function when data of each element of the set is input to the first function and the second function by the set element input unit, or measures a similarity between first output data from the first approximation function and second output data from the second approximation function when data of each element of the set is input to the first approximation function and the second approximation function by the set element input unit; an output unit that outputs a result related to the similarity measured by the similarity measurement unit; A measuring device having:

2. 2. The measuring device according to claim 1, a domain conversion unit that converts the domain of the data in the first input range or the data in the second input range when the data in the first input range and the data in the second input range are input by the input unit and the data in the first input range and the data in the second input range are not identical, thereby aligning the first input range and the second input range to data in the specified input range.

3. 2. The measuring device according to claim 1, an input range acquisition unit that, when the data in the first input range and the data in the second input range are not input by the input unit, acquires data in a predetermined first input range from the first machine learning model and acquires data in a predetermined second input range from the second machine learning model; the input range determination unit determines whether the data of the predetermined first input range and the data of the predetermined second input range acquired by the input range acquisition unit are identical, and if it is determined that they are identical, it regards the data of the predetermined first input range and the data of the predetermined second input range as data of the same predetermined input range; Measuring equipment.

4. 4. The measuring device according to claim 3, a domain conversion unit that, when the data in the predetermined first input range and the data in the predetermined second input range are not identical, converts the domain of the data in the predetermined first input range or the data in the predetermined second input range to align the first input range and the second input range to form the data in the predetermined input range.

5. A measuring device according to any one of claims 1 to 4; a communication terminal capable of transmitting the first input range data and the second input range data to the measurement device; A communication system having:

6. A measurement method performed by a measurement device that measures changes in the behavior of a machine learning model, The measuring device is an input range determination process in which, when data in a first input range and data in a second input range are input, it is determined whether the data in the first input range and the data in the second input range are identical, and, when it is determined that they are identical, the data in the first input range and the data in the second input range are set as data in the same predetermined input range; a set setting process for setting the data in the predetermined input range itself as data in the set for input, or for setting values ​​sampled from the data in the predetermined input range as data in the set for input; a set element input process for inputting data of each element of the set into a first function related to a first machine learning model and a second function related to a second machine learning model, or for inputting data of each element of the set into a first approximation function that approximates a decision tree or rule simpler than the first machine learning model and a second approximation function that approximates a decision tree or rule simpler than the second machine learning model; a similarity measurement process for measuring a similarity between first output data from the first function and second output data from the second function when data of each element of the set is input to the first function and the second function, respectively, by the set element input process, or for measuring a similarity between first output data from the first approximation function and second output data from the second approximation function when data of each element of the set is input to the first approximation function and the second approximation function, respectively, by the set element input process; an output process for outputting a result related to the similarity measured by the similarity measurement process; How to perform the measurement.

7. A program causing a computer to execute the method according to claim 6.

Images