Machine learning device, machine learning method, and program

Abstract

A machine learning device according to one embodiment of the present invention is provided with a storage unit and a processor. The storage unit stores answer data, including parameters presented in relation to measurement of a psychometric function and users' evaluation values for the parameters. The processor includes a logistic regression unit, a parameter selection unit, and a parameter presentation unit. The logistic regression unit repeatedly inputs parameters and evaluation values of the answer data to a logistic regression model and thereby trains the logistic regression model through machine learning to generate a trained model corresponding to the psychometric function. The parameter selection unit selects parameters to be presented, through active learning based on evaluation values of the answer data and the trained model. The parameter presentation unit presents the selected parameters to a subject.

Description

Machine learning device, machine learning method, and program

[0001] One aspect of the present invention relates to a machine learning device, a machine learning method, and a program.

[0002] Psychometric functions are often used in fields such as psychology and affective machine learning as a way to measure how people feel about something, such as evaluations using the five senses. Psychometric functions are functions that express the relationship between the physical quantity of a stimulus given to a person and the reaction to that stimulus, and can be measured. For example, they can quantitatively evaluate whether a temperature is perceived as hot, estimate emotions and likeability from facial expressions, gestures, and voices, or how people perceive the aroma, taste, and deliciousness of food and drink.

[0003] For example, when one wishes to evaluate how much a facial expression intensity parameter is perceived as happy, a psychometric function is measured to obtain the psychometric function itself. Furthermore, a psychometric function is said to take the form of a growth curve, or S-curve. For example, when one wishes to evaluate the minimum facial expression intensity perceived as a smile, a psychometric function is measured to find the inflection point of the psychometric function. Alternatively, when one wishes to know which of the eye and mouth intensity parameters influences how much a face is perceived as happy, a psychometric function is measured to evaluate the influence of each dimension in a multidimensional psychometric function.

[0004] Yasuhisa Nakano, “Psychophysical measurement method”, Vision 7, pp. 17-27, 1995.

[0005] The constant method is known as a representative method for measuring psychometric functions (see, for example, Non-Patent Document 1). This method presents stimuli to the subject while randomly changing their intensity, and is considered to be less susceptible to response errors due to subject predictions, etc., compared to the adjustment method (in which the subject adjusts the intensity of the stimuli themselves) and the method of limits (in which the experimenter presents stimuli to the subject while changing the intensity by a fixed amount).

[0006] However, the constant method tends to require a large number of inquiries from subjects, and is particularly difficult to apply in practice to measure multidimensional psychometric functions, since the number of inquiries increases further due to the so-called curse of dimensionality.

[0007] The present invention has been made in light of the above circumstances, and aims to provide a technique that makes it possible to obtain a highly accurate psychometric function with a minimum number of inquiries.

[0008] A machine learning device according to one aspect of the present invention includes a memory unit and a processor. The memory unit stores response data including presented parameters related to measurement of a psychometric function and a user's evaluation value for the parameters. The processor includes a logistic regression unit, a parameter selection unit, and a parameter presentation unit. The logistic regression unit repeatedly provides the parameters and evaluation value of the response data to a logistic regression model for machine learning, thereby generating a trained model corresponding to the psychometric function. The parameter selection unit selects parameters to be presented by active learning based on the evaluation value of the response data and the trained model. The parameter presentation unit presents the selected parameters to the subject.

[0009] According to one aspect of the present invention, it is possible to provide a technique that can obtain highly accurate psychometric functions with a minimum number of inquiries.

[0010] FIG. 1 is a functional block diagram showing an example of a machine learning device according to an embodiment. FIG. 2 is a diagram showing an example of data stored in the answer result database 31a. FIG. 3 is a diagram showing an example of the flow of data in the machine learning device 10. FIG. 4 is a graph showing an example of a curve drawn by a psychometric function. FIG. 5 is a diagram for explaining a psychometric function related to facial expression. FIG. 6 is a diagram showing that the psychometric function changes each time regression is repeated. FIG. 7 is a functional block diagram showing an example of an existing device that performs learning using a constant method. FIG. 8 is a block diagram showing an example of the hardware configuration of the machine learning device 10.

[0011] Emotions are how people feel about certain events, and are heavily dependent on factors such as individual personality, experience, and memory. Therefore, even when given the same information under the same circumstances, the information received varies significantly from person to person. In particular, in interpersonal communication, the way the other person perceives something varies greatly, and it is possible that the information will be received in a way that the sender did not anticipate. Some people tend to interpret the other person's emotions more negatively than necessary, and may even feel uncomfortable with communication itself. Psychometric functions can be useful in smoothing communication between sender and receiver.

[0012] 1 is a functional block diagram showing an example of a machine learning device according to an embodiment. The machine learning device 10 is a computer and includes an input unit 40, a processor 20, a storage unit 30, and an output unit 50.

[0013] The input unit 40 includes a user response input unit 41. The user response input unit 41 accepts input of a response result from a user (subject) in response to a presented parameter (stimuli) related to the measurement of a psychometric function. The response result is given as an evaluation value for the parameter. The presented parameter and the user's evaluation value for this parameter are stored as a set in the storage unit 30.

[0014] The storage unit 30 stores a database 31 and a model 32. The database 31 includes an answer result database 31a. The answer result database 31a is a database that holds answer data including parameters and evaluation values.

[0015] FIG. 2 is a diagram showing an example of data stored in the answer result database 31a. For example, consider optimizing a cookie recipe based on a user's answers. When a user scores a certain recipe (e.g., the amount and ratio of ingredients), another recipe is presented and the user is prompted to score it again. By repeating this process, a recipe with the highest user rating can be obtained. In such a case, the answer result database 31a in FIG. 2 stores parameters that differ for each recipe and the user's answer results, given on a scale of, for example, 10 points, for each attempt.

[0016] 1, the description will be continued. The model 32 in the storage unit 30 is a trained model generated by the machine learning device 10.

[0017] The processor 20 includes a logistic regression unit 200, a parameter selection unit 201, and a parameter presentation unit 202 as processing functions according to the embodiment.

[0018] The logistic regression unit 200 repeatedly provides the parameters of the response data and the evaluation values ​​stored in the response result database 31a to an untrained neural network for machine learning. In this embodiment, a logistic regression model is assumed as the neural network. A trained model corresponding to the psychometric function is generated by machine learning. The generated model is stored in the storage unit 30 as a model 32.

[0019] The logistic regression unit 200 selectively uses either logistic regression or Bayesian logistic regression depending on the user's response error for the presented parameters. Psychometric functions are obtained as regression results, and the influence of each dimension is obtained as a regression coefficient.

[0020] The parameter selection unit 201 uses active learning to select parameters to be presented based on the evaluation values ​​of the response data and the trained model 32. In this embodiment, the psychometric function is assumed to be a logistic distribution. The parameter selection unit 201 uses active learning (reference document [1]) to select questions (parameters) based on past responses and presents them to the subject. The model is trained based on the responses (evaluation values) obtained here.

[0021] The parameters to be presented to the subject can be searched for by applying, for example, uncertainty sampling, disagreement sampling, or information density algorithms (Reference [1]). The parameter presenting unit 202 presents the selected parameters to the subject.

[0022] FIG. 3 is a diagram showing an example of the flow of data in the machine learning device 10. Answer results input from the user answer input unit 41 are stored in the answer result database 31a of the storage unit 30. The logistic regression unit 200 acquires answer data, generates a trained model 32 through machine learning, and stores it in the storage unit 30. The parameter selection unit 201 acquires the evaluation value of the answer data and the trained model 32, and selects parameters to be presented to the user. The selected parameters are displayed by the parameter presentation unit 202, for example, on a browser screen of a display (not shown), prompting the user to provide an answer. In response, the answer results are input again, and the logistic regression unit 200 repeatedly fine-tunes the model 32.

[0023] 4 is a graph showing an example of a curve drawn by a psychometric function. When the parameter value is plotted on the horizontal axis and the evaluation value on the vertical axis, the psychometric function draws a growth curve (S-shaped curve) with an inflection point near the center of the domain.

[0024] Next, the machine learning of the embodiment will be described in detail using mathematical expressions. In the following, an example will be described in which a forced choice between positive and negative is made when presented with a parameter representing the intensity of a human facial expression that continuously changes from a smiling face to a crying face.

[0025] Assuming that the psychometric function is a logistic distribution, a and b in equation (1) are estimated. a and b can be calculated using Newton's method or the like so that the log likelihood shown in equations (2) and (3) is maximized. In equations (2) and (3), the i-th response data is (x ｉ , y ｉ ) where x ｉ is the presentation parameter, y ｉ represents the answer.

[0026]

[0027] FIG. 5 is a diagram for explaining psychometric functions related to facial expressions. The horizontal axis of the graph represents parameter values, and the vertical axis represents response probability. The domain of parameter values ​​is set to -0.1 to +0.1, with the maximum negative value (-0.1) corresponding to the maximum crying face intensity parameter and the maximum positive value (0.1) corresponding to the maximum smiling face intensity parameter. The response probability is expressed as follows: a lower value indicates a more negative response, and a higher value indicates a more positive response. Here, a one-dimensional example is shown, with values ​​representing a forced choice between positive and negative.

[0028] The prior knowledge that the most crying face is negative (x = -0.1, y = 0) and the most smiling face is positive (x = 0.1, y = 1) is input into the model. The asterisks (*) on both ends of the graph correspond to these data. The obtained regression results are expressed, for example, by Equation (4).

[0029]

[0030] Equation (4) shows that when the horizontal axis changes by 1, the probability of responding positive is e ３７．９ This can be interpreted as a factor of 2 (easily influenced). Also, when the horizontal axis is 0, the probability of responding positively is e ０ / (1+e ０ ) can be interpreted as 50% (positivity).

[0031] Next, the x with the highest uncertainty is searched for from the regression results using parameter search (e.g., uncertainty sampling). For example, equation (5) can be used. However, (6) represents the class with the highest classification probability, either positive (i.e., 1) or negative (i.e., 0). For example, (7) becomes (8).

[0032]

[0033] As a result of the search, 0.000 is selected as the parameter to be presented. In other words, 0.000 is selected this time as the most uncertain parameter, and the facial expression corresponding to this parameter (the facial expression in the middle) is presented to the subject. If the subject answers negative (0) at this time, the regression result based on this answer is expressed by equation (9).

[0034]

[0035] Figure 6 shows how the psychometric function changes with each iteration of the regression. The psychometric function of Equation (9) is shown in Figure 6(a). The asterisk (*) at 0.000 on the horizontal axis indicates the current response, and the regression results include past responses (*) on both ends of the curve.

[0036] Furthermore, suppose that a parameter search is performed again based on the curve in Figure 6(a), and the most uncertain value of 0.050 is queried, and the subject answers negative (0). The regression result based on this answer is expressed by equation (10). The regression result at this time is shown in Figure 6(b).

[0037]

[0038] By repeating this type of learning (active learning), it becomes possible to respond to any input. Figure 6(c) shows the regression results after one more stage of learning has been carried out from Figure 6(b).

[0039] As described above, according to the embodiment, by learning logistic regression by active learning based on binary responses from subjects, it is possible to obtain a psychometric function with a small number of trials. Furthermore, when the response error is large enough that ordinary logistic regression cannot withstand it, machine learning with enhanced noise tolerance can be realized by using Bayesian logistic regression, which is resistant to noise.

[0040] For comparison, Figure 7 is a functional block diagram showing an example of an existing device that performs learning using the constant method. In the existing device, the parameters to be given to the subject were randomly selected by a parameter random selection unit 300. In other words, the regression was performed using data that was, so to speak, blindly collected without applying an active learning method, which tended to result in a large number of trials.

[0041] In contrast, in the embodiment, parameters are selected by referring to a trained model obtained by logistic regression, and recursive regression is performed based on the response, so that a highly accurate psychometric function can be obtained even with as few inquiries as possible from the user. As a result, according to the embodiment, a highly accurate psychometric function can be obtained with a minimum number of inquiries.

[0042] The functions of the machine learning device 10 can be realized by installing a program on a computer. For example, the functions of the machine learning device 10 can be implemented by having a computer execute a program provided as packaged software or online software. The computer is not limited to a so-called server computer, and may be a desktop or notebook personal computer. Portable terminals such as smartphones and tablets are also included in the category of computers. Furthermore, computer resources virtualized on the cloud can also function as the machine learning device 10.

[0043] 8 is a block diagram showing an example of the hardware configuration of the machine learning device 10. As shown in FIG. 6, the machine learning device 10 includes a CPU (Central Processing Unit) 20A, a bridge circuit 102, memory 60, a GPU (Graphics Processing Unit) 20B connected to a display 51, a BIOS-ROM 70, storage 80, a USB connector 90, and an input unit 40.

[0044] The storage 80 is a non-volatile storage medium (block device), such as a hard disk drive (HDD) or a solid state drive (SSD). The storage 80 stores basic programs such as an operating system (OS) 62 and device drivers, as well as a program 61 for implementing the functions of the machine learning device 10.

[0045] The memory 60 includes a ROM (Read Only Memory) and a RAM (Random Access Memory). The CPU 20A and the GPU 20B are arithmetic elements related to the processor 20. The CPU 20A mainly controls the machine learning device 10. The GPU 20B mainly performs calculations related to image processing (such as multiply-accumulate operations) at high speed. The CPU 20A executes a BIOS (basic input / output system) stored in the BIOS-ROM 70. The CPU 20A also loads a program 61 from the storage 80 into the memory 60 and executes it. The same is true for the GPU 20B.

[0046] The bridge circuit 102 relays data transmission between the CPU 20A and the GPU 20B and each component. For example, the bridge circuit 102 is interposed between the CPU 20A and the GPU 20B and hardware devices connected to a PCI (Peripheral Component Interconnect) bus or a PCIe bus (not shown), relaying communication between them.

[0047] The USB connector 90 connects to a USB device or the like. For example, the program 61 may be installed in the machine learning device 10 via a USB device. The input unit 40 connects the machine learning device 10 to a communication network such as a local area network (LAN) or a wide area network (WAN).

[0048] The program 61 and various data may be stored in a removable storage medium other than the storage 80, and may be read by the CPU 20A from a disk drive or the like. Alternatively, the program 61 and various data may be stored in another computer connected via a communication network, and may be read by the CPU 20A via the input unit 40.

[0049] In short, this invention is not limited to the above-described embodiments, and various modifications can be made in the implementation stage without departing from the spirit of the invention. Furthermore, the embodiments may be implemented in appropriate combinations, in which case the combined effects can be obtained. Furthermore, the above-described embodiments include various inventions, and various inventions can be extracted by combining selected elements from the disclosed elements. For example, if the problem can be solved and the desired effect can be obtained even if some elements are deleted from all elements shown in the embodiments, the configuration from which these elements are deleted can be extracted as an invention.

[0050] <References> [1] Kumar P, Gupta A: “Active Learning Query Strategies for Classification, Regression, and Clustering: A Survey”, Journal of computer science and technology, 2020.

[0051] 10...machine learning device 20...processor 20A...CPU 30...memory unit 31...database 31a...answer result database 32...trained model 40...input unit 41...user answer input unit 50...output unit 51...display 60...memory 61...program 80...storage 90...USB connector 102...bridge circuit 200...logistic regression unit 201...parameter selection unit 202...parameter presentation unit 300...parameter random selection unit.

Claims

1. A machine learning device comprising a memory unit and a processor, wherein the memory unit stores answer data including parameters presented in relation to the measurement of a psychometric function and a user's evaluation value of the parameters, and the processor comprises: a logistic regression unit that repeatedly supplies the parameters and evaluation value of the answer data to a logistic regression model for machine learning to generate a trained model corresponding to the psychometric function; a parameter selection unit that selects parameters to be presented by active learning based on the evaluation value of the answer data and the trained model; and a parameter presentation unit that presents the selected parameters to a subject.

2. The machine learning device according to claim 1, wherein the logistic regression unit selectively uses either logistic regression or Bayesian logistic regression depending on the response error for the presented parameters.

3. A machine learning method by a computer having a memory unit that stores answer data including parameters presented in relation to the measurement of a psychometric function and a user's evaluation value of the parameters, and a processor, comprising: a step in which the processor repeatedly supplies the parameters and evaluation value of the answer data to a logistic regression model for machine learning to generate a trained model corresponding to the psychometric function; a step in which the processor selects parameters to be presented by active learning based on the evaluation value of the answer data and the trained model; and a step in which the processor presents the selected parameters to a subject.

4. A program comprising instructions for causing a computer to function as the machine learning device according to claim 1 or 2.

Images