Emotional estimation device

Abstract

To provide an emotion estimation apparatus which implements a configuration improved in accuracy of estimating emotion so as to bring greater satisfaction to a target person while suppressing reduction in degree of freedom of design for hardware resources.SOLUTION: An emotion estimation apparatus 1 includes a control unit 20 having an emotion estimation model 30 for estimating emotion based on biometric information Bi, to estimate emotion of a target person U based on biometric information Bi of the target person U. The emotion estimation model 30 includes a secondary biometric information distillation model which is generated by using primary knowledge information Ki1. The control apparatus 20 is configured to: acquire meta-learning X biometric information BiMx and meta-learning X self emotion evaluation information EiMx of a target person Ux in an environment which is different from an environment in which the biometric information Bi used in generating a primary model 100 was obtained; use a secondary biometric information distillation X model 31x adjusted so that the characteristics of the target person Ux is reflected; and output secondary emotion information Ei2 in which the different environment or different characteristics of the target person Ux are reflected, based on the biometric information Bix of the target person Ux.SELECTED DRAWING: Figure 6

Description

【Technical Field】 【0001】 The present invention relates to an emotion estimation device that estimates the emotion of a target person based on the biological information of the target person by using an emotion estimation model. 【Background Art】 【0002】 There is known an emotion estimation device that estimates the emotion of a target person based on the biological information of the target person by using an emotion estimation model. For example, Non-Patent Document 1 discloses a method of estimating the emotion of a target person based on the biological information of the target person by using an emotion estimation model learned (knowledge distillation) based on biological information and knowledge information of a teacher model. 【0003】 The emotion estimation model possessed by the emotion estimation device is created by knowledge distillation using the output of a teacher model that estimates emotion based on a plurality of biological information. The output of the teacher model inherits information for estimating emotion based on the biological information of the target person obtained by the teacher model through learning based on a plurality of types of biological information. That is, the emotion estimation model inherits information that is not included in the biological information used for learning by using the output of the teacher model. Therefore, by using the emotion estimation model, the emotion estimation device can obtain an estimation accuracy equivalent to that of the teacher model based on a smaller number of types of biological information than the number of types of biological information required for estimating the emotion of the target person in the teacher model. 【Prior Art Documents】 【Non-Patent Documents】 【0004】 【Non-Patent Document 1】 Tsubasa Nishihara, et al. 7 persons, "Driver Emotion Estimation System Based on Multimodal Biological Signals", June 2021, Proceedings of the Robotics and Mechatronics Conference 2021, 2P3-J17 【Summary of the Invention】 【Problems to be Solved by the Invention】 【0005】 As described in Non-Patent Document 1, by using an emotion estimation model generated by knowledge distillation based on the subject's biometric information to estimate the subject's emotions, biometric information that can be easily obtained from the subject can be used to estimate emotions. Therefore, while improving the convenience of obtaining biometric information when estimating the subject's emotions, it is possible to improve the accuracy of emotion estimation compared to using an emotion estimation model that is not generated by knowledge distillation. 【0006】 Furthermore, by using the emotion estimation model described in Non-Patent Literature 1 above to estimate the subject's emotions based on the biometric information, the amount of biometric data used for emotion estimation can be reduced compared to using an emotion estimation model that is not generated by knowledge distillation. Therefore, the computational load on the emotion estimation device can be reduced, and the design flexibility of the hardware resources in the emotion estimation device can be improved. 【0007】 Therefore, by using the emotion estimation model described in Non-Patent Literature 1 above to estimate the subject's emotions based on the biometric information, the emotion estimation device can achieve both improved accuracy in emotion estimation and increased design flexibility for hardware resources compared to using an emotion estimation model that was not created through knowledge distillation. 【0008】 However, in the course of developing the emotion estimation device, the inventors found that there is room to improve the subject's sense of satisfaction with the emotion estimation in certain situations or for certain subjects, and therefore further improvements in the accuracy of emotion estimation are possible. 【0009】 Therefore, there is a need for a configuration in the emotion estimation device that can further enhance the subject's sense of satisfaction and improve the accuracy of emotion estimation while suppressing a decrease in the design flexibility of hardware resources. 【0010】 The present invention aims to realize a configuration in an emotion estimation device that uses the emotion estimation model to estimate the emotions of a subject, which improves the accuracy of emotion estimation by increasing the subject's sense of satisfaction while suppressing a decrease in the design flexibility of hardware resources. [Means for solving the problem] 【0011】 The inventors diligently studied an emotion estimation device that improves the accuracy of emotion estimation by increasing the subject's sense of satisfaction while suppressing a decrease in the design flexibility of hardware resources. 【0012】 To further improve the accuracy of emotion estimation, one could consider increasing the types and amount of biometric data used. However, simply increasing the types and amount of biometric data tends to reduce the design flexibility of hardware resources. 【0013】 Therefore, the inventors attempted to investigate the accuracy of emotion estimation. As a result of their investigation, the inventors found that the accuracy of emotion estimation is determined by the degree of agreement between the estimated emotion of the subject, estimated using the subject's measured biosignals, and the cognitive emotion recognized by the subject's own brain when the biosignals were measured. 【0014】 Furthermore, while investigating ways to suppress the discrepancy between estimated emotions and cognitive emotions, the inventors found that instead of increasing the types or amount of biometric data, they could enhance the subject's sense of acceptance (understanding) by using a limited number of biometric data types and a limited amount of data for a particular scene or subject, and by tuning (adjusting) an emotion estimation model created through knowledge distillation to reduce input. As a result, the inventors found that the discrepancy between estimated emotions and cognitive emotions is suppressed, thereby improving the accuracy of emotion estimation. 【0015】 Moreover, since this is achieved by using a limited number of biometric data types and a small amount of data for a particular scene or subject, and by tuning a biometric information distillation model that reduces the input created by knowledge distillation, it is possible to suppress a decrease in the design flexibility of the hardware resources in the emotion estimation device. 【0016】 An emotion estimation device according to one embodiment of the present invention comprises a storage unit for storing biological information, and a control unit that has an emotion estimation model used for estimating emotions based on the biological information, acquires the biological information of a subject and stores it in the storage unit, estimates the emotions of the subject based on the biological information of the subject stored in the storage unit using the emotion estimation model, and outputs information relating to the estimated emotions as emotion information. If a model used to estimate emotions based on multiple types of the aforementioned bio-information is defined as a first-order model, and a model created using n-1-order bio-information distillation models, such as a second-order bio-information distillation model created using the first-order knowledge information obtained by the first-order model, or a third-order bio-information distillation model created using the second-order knowledge information obtained by the second-order bio-information distillation model, is defined as an n-order bio-information distillation model (where n is 2 or an integer greater than 2), then the emotion estimation model includes the n-order bio-information distillation model used to estimate the subject's emotions based on the number of types of bio-information that are included in the types of bio-information necessary for estimating emotions in the first-order model and are less than the number of types of bio-information necessary for estimating emotions in the first-order model, then the control unit uses the n-order bio-information distillation model to estimate the subject's emotions based on the subject's bio-information that is included in the types of bio-information necessary for estimating emotions in the first-order model and is less than the number of types of bio-information necessary for estimating emotions in the first-order model, then outputs the estimated emotion information as emotion information. 【0017】 The control unit acquires biometric information of the same type as the biometric information input to the nth-order biometric information distillation model, measured in an environment different from the environment in which the biometric information used to create the first-order model was acquired, and self-emotion evaluation information in which the provider of the biometric information evaluated their own emotions at the time the biometric information was measured, or biometric information of the same type as the biometric information input to the nth-order biometric information distillation model, measured by a provider different from the provider of the biometric information used to create the n-1th-order biometric information distillation model, and self-emotion evaluation information in which the provider of the biometric information evaluated their own emotions at the time the biometric information was measured. The control unit uses the nth-order biometric information distillation model, which has been adjusted to reflect the characteristics of the different environment or the different provider based on the acquired biometric information and self-emotion evaluation information, to output emotion information that better reflects the characteristics of the different environment or the different subject, based on the subject's biometric information which is included in the types of biometric information necessary for estimating emotions in the first-order model and has a number of types less than the number of types of biometric information necessary for estimating emotions in the first-order model. 【0018】 As described above, the biometric information used when adjusting the nth-order biometric information distillation model is of the same type as the biometric information input to the nth-order biometric information distillation model, and is biometric information acquired under different conditions than the environment in which the primary biometric information used when creating the primary model was acquired, or biometric information of a different subject. Furthermore, the self-emotion evaluation information used to adjust the nth-order biometric information distillation model is self-emotion evaluation information obtained when acquiring biometric information of the same type as the biometric information input to the nth-order biometric information distillation model, self-emotion evaluation information obtained under different conditions than the environment in which the primary biometric information used when creating the primary biometric information distillation model was acquired, or self-emotion evaluation information of a different subject. In this way, the nth-order biometric information distillation model is adjusted using biometric information of the same type as the biometric information input to the nth-order biometric information distillation model obtained through learning using the knowledge information of the teacher model, and self-emotion evaluation information obtained when the biometric information was acquired. As a result, since there is no need to increase the types of biometric information when adjusting the nth-order biometric information distillation model, the decrease in the design degree of freedom of the hardware resources in the emotion estimation device can be suppressed. 【0019】 Moreover, the outputted estimated emotion information is information about emotions estimated using the nth-order bio-information distillation model, which is adjusted to reflect the characteristics of different environments or different subjects, based on the bio-information acquired in an environment different from the environment in which the primary bio-information used in creating the primary model was acquired, and the self-emotion evaluation information which is information related to the emotions perceived by the subject themselves, or the bio-information of different subjects and the self-emotion evaluation information. Therefore, the characteristics of different environments or different subjects are better reflected. In this way, the subject's emotions are estimated using the nth-order bio-information distillation model, which takes into account the relationship between the subject's bio-information and the cognitive emotions perceived by the subject in their brain. As a result, the discrepancy between the estimated bio-emotions estimated based on bio-information and the cognitive emotions can be suppressed, the subject's sense of satisfaction can be further enhanced, and the accuracy of emotion estimation can be further improved. 【0020】 Therefore, in an emotion estimation device that estimates the emotion of a target person using an emotion estimation model, it is possible to realize a configuration that can improve the accuracy of emotion estimation by enhancing the satisfaction of the target person while suppressing a decrease in the design freedom of hardware resources. 【0021】 From another perspective, it is preferable that the emotion estimation device of the present invention includes the following configuration. The n-th biological information distillation model is adjusted based on the biological information and the self-emotion evaluation information obtained under the different environments or obtained from the different target persons, and uses the initial parameters of the n-th biological information distillation model to be adjusted so that the characteristics of the different environments or the different target persons are reflected. 【0022】 As described above, the n-th biological information distillation model is adjusted so as to reflect the characteristics of a new target person by using an existing n-th biological information distillation model that has already been adjusted based on the biological information and the self-emotion evaluation information of other target persons. Therefore, since it is not necessary to increase the type and amount of biological information in improving the accuracy of emotion estimation, it is possible to suppress a decrease in the design freedom of hardware resources in the emotion estimation device. Furthermore, since an existing adjusted n-th biological information distillation model is used when estimating emotions, it is possible to realize a configuration that can improve the accuracy of emotion estimation by enhancing the satisfaction of the target person. 【0023】 The technical terms used in this specification are used only for the purpose of defining specific embodiments, and are not intended to limit the invention by these technical terms. 【0024】 As used in this specification, "and / or" includes all combinations of one or more of the related listed components. 【0025】 As used herein, the use of "including", "comprising", "having" and their variations identify the presence of the stated features, steps, operations, elements, components, and / or their equivalents, but can include one or more of steps, actions, elements, components, and / or groups thereof. 【0026】 As used herein, "attached", "connected", "coupled", and / or their equivalents are used in a broad sense and encompass both "direct and indirect" attachment, connection, and coupling. Further, "connected" and "coupled" are not limited to physical or mechanical connections or couplings, and can include direct or indirect connections or electrical connections or couplings. 【0027】 Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. 【0028】 Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the relevant art and this disclosure, and should not be interpreted in an idealized or overly formal sense unless expressly defined herein. 【0029】 In the description of the present invention, it is understood that several techniques and processes are disclosed. Each of these has individual benefits and can also be used in combination with one or more, or in some cases all, of the other disclosed techniques. 【0030】 Therefore, for clarity, the description of the present invention refrains from repeating all possible combinations of the individual steps unnecessarily. However, this specification and the claims should be read with the understanding that all such combinations are within the scope of the present invention. 【0031】 This specification describes embodiments of the emotion estimation device according to the present invention. 【0032】 The following description includes numerous specific examples to provide a complete understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention can be carried out without these specific examples. 【0033】 Therefore, the following disclosures should be considered illustrative examples of the present invention and are not intended to limit the present invention to any specific embodiments shown in the following drawings or description. 【0034】 [Target audience] In this specification, "subject" refers to a person whose emotions are estimated by an emotion estimation device. The subject's biological information is acquired by the emotion estimation device. Furthermore, the subject's emotions are estimated by the emotion estimation device based on the acquired biological information. 【0035】 [environment] In this specification, "environment" means the surrounding conditions in which the subject and subject are placed, the work they are doing, and any other circumstances that provide information and stimuli to the subject and subject from the outside. The environment includes, for example, a state in which the subject is moving in a lean vehicle, a state in which the subject is moving on foot, a state in which the subject is looking at a photograph of an animal, and each of these states includes indoors or outdoors. 【0036】 [Provider] In this specification, a provider is a person whose biometric information is measured for use in training a biometric information distillation model used to estimate emotions. The provider also has their biometric information measured in response to specific external events and provides information about emotions they perceive themselves (self-emotion assessment information). For each biometric information in response to an external event, the provider provides information about emotions through self-reporting. The biometric information for learning and the information about emotions constitute training data. 【0037】 [subject] In this specification, a subject is a person whose biometric information is measured for use as training data for training an emotion estimation model used to estimate emotions. The subject has their biometric information measured in response to specific external events and provides emotion information, which is information about emotions, through self-reporting. The subject provides emotion information through self-reporting for each biometric information in response to external events. The subject is a provider of training data. 【0038】 [study] In this specification, learning or machine learning means the process of adjusting the parameters of a machine learning model (program) in a machine (processor) so that the output is correct for the input. In this embodiment, the primary model, secondary bio-information distillation model, and tertiary bio-information distillation model to be learned are subjected to at least supervised learning, which is performed on training data that combines input data and the correct answers for the input data, or unsupervised learning which determines whether an answer is correct or not using only the input data. 【0039】 [Biometric Information] In this specification, biological information refers to various physiological and anatomical information emitted by a subject or test subject. Biological information means all life activities occurring in the biological body of the subject or test subject. Biological information includes, for example, information on the subject's heart rate, heart sounds, heart waveform, heart cycle, changes in heart rate, blood pressure, pulse wave, triaxial acceleration, body surface temperature, electroencephalogram, respiratory rate, pupil state, electromyography, blood components, exhalation, exhalation volume, exhalation components, etc. Biological information is measured by wearable sensors worn by the subject, such as heart rate sensors, acceleration sensors, and temperature sensors. 【0040】 [Biometric information for learning] In this specification, "learning biometric information" refers to the biometric information of a subject used for learning an n-1 order biometric distillation model, such as a first-order model, a second-order biometric distillation model, and a third-order biometric distillation model. Learning biometric information used for learning a first-order model is referred to as first-order learning biometric information, learning biometric information used for learning a second-order biometric distillation model is referred to as second-order learning biometric information, and so on, with learning biometric information used for learning an n-1 order biometric distillation model being referred to as n-1 order learning biometric information. Furthermore, learning biometric information used for meta-learning is referred to as meta-learning biometric information. Learning biometric information includes, for example, information regarding the subject's heart rate, heart sounds, heart rate waveform, heart rate cycle, changes in heart rate, blood pressure, pulse wave, triaxial acceleration, body surface temperature, electroencephalogram, respiratory rate, pupillary state, electromyography, blood components, exhalation, exhalation volume, exhalation components, etc. Learning biometric information is measured by wearable sensors worn by the subject, such as heart rate sensors, acceleration sensors, and temperature sensors. 【0041】 [Biometric information related to heart rate] In this specification, biometric information relating to heart rate refers to biometric information including the subject's heart rate, heart cycle, changes in heart rate, blood pressure, electrocardiogram signal, pulse wave, etc. Biometric information relating to heart rate is measured by a heart rate sensor provided on a wearable sensor or the like, which is a biometric information acquisition unit worn by the subject. 【0042】 [Emotional information] In this specification, emotional information means information indicating a state of feeling emotions such as joy, anger, sadness, etc. The emotional information includes information indicating a state in which the subject is feeling happy, relaxed, angry, and sad. The emotional information includes information regarding the probability of the subject's emotion being estimated based on acquired biometric information. The emotional information may also include information regarding the degree of the emotion the subject is feeling. One piece of emotional information may contain information about multiple emotions. The emotional information is created, for example, by a control unit located in a mobile device held by the subject or in a server that acquired the biometric information. 【0043】 [Self-emotional assessment information] In this specification, self-emotional evaluation information refers to information related to the emotions perceived by a subject in response to an external event at the time the biological information was measured. Self-emotional evaluation information is, for example, emotional information such as joy, anger, sadness, and pleasure perceived by the subject in response to visual information such as a landscape, which is an external event. Self-emotional evaluation information used for learning a first-order model is referred to as first-order learning self-emotional evaluation information, self-emotional evaluation information used for learning a second-order biological information distillation model is referred to as second-order learning self-emotional evaluation information, and so on, with self-emotional evaluation information used for learning an n-1-order biological information distillation model being referred to as n-1-order learning self-emotional evaluation information. Furthermore, self-emotional evaluation information used for meta-learning is referred to as meta-learning self-emotional evaluation information. 【0044】 [Characteristics of the target group] In this specification, subject characteristics refer to the relationship between biometric information and emotional information in a subject. The subject characteristics include the content of biometric information when the subject is experiencing any given emotion. The subject characteristics differ from subject to subject. Therefore, biometric information in the same environment differs from subject to subject. Furthermore, the emotions a subject experiences when their biometric information is the same differ from subject to subject. Also, the biometric information when experiencing the same emotion differs from subject to subject. In other words, the relationship between biometric information, emotional information, and the environment differs from subject to subject. 【0045】 [MAML] In this specification, MAML (Model-Agnostic Meta-Learning) means meta-learning that is independent of the machine learning model. Meta-learning means learning a machine learning method for a new task through multiple other related tasks. MAML updates the parameters of the machine learning model to adapt it to a new task. MAML updates the parameters of the machine learning model through learning by an inner loop and learning by an outer loop. The learning by the inner loop is learning that updates the parameters of the machine learning model assigned to each task based on multiple learning tasks (in this embodiment, biometric information for meta-learning and self-emotion evaluation information for meta-learning from multiple subjects). The learning by the outer loop is learning that updates the parameters of the machine learning model before learning by the inner loop using error propagation or the like, based on each of the parameters of the machine learning model updated in the inner loop. The parameters updated in the learning by the outer loop are the initial values ​​of the parameters of the machine learning model for a new task different from the multiple tasks (in this embodiment, biometric information and self-emotion evaluation information in an environment different from the environment in which the biometric information of the multiple subjects was acquired, or biometric information and self-emotion evaluation information of a subject who is a provider of biometric information different from the subjects). MAML can learn with less data than conventional machine learning by applying the learning results from the inner loop to the learning from the outer loop. 【0046】 [Fine tuning] In this specification, fine-tuning means updating the initial parameters of a machine learning model with initial parameters set in MAML based on a new task (in this embodiment, meta-learning biometric information and meta-learning self-emotional assessment information from an environment different from the environment in which the biometric information of the multiple subjects was acquired, or meta-learning biometric information and meta-learning self-emotional assessment information from a subject who is a provider of biometric information different from the subjects). The machine learning model is learned through fine-tuning to have parameters specialized for emotion estimation in a specific environment or for a specific subject. The initial values ​​of the machine learning model's parameters are adjusted through fine-tuning to parameters specialized for the new task. 【0047】 [Model] In this specification, "model" refers to a specific mathematical formula, function, or method used in machine learning. A model is a mathematical formula that expresses the relationship between input and output. The model transforms input data according to the formula. Creating the model means updating the parameters of each node in the model through learning. Examples of such models include regression models that predict values ​​and classification models that classify data. 【0048】 [Neural Network Model] In this specification, a neural network model means a mathematical model in which multiple processing units that linearly transform inputs are connected to each other. The neural network model has an input layer, at least one hidden layer, an output layer, a function connecting the input layer and the hidden layer, and a function connecting the hidden layer and the output layer. The input layer, the hidden layer, and the output layer are multiple variables into which data is input. The neural network model sequentially transforms the data input to the input layer using the function and outputs it. When training data consisting of combinations of data and the correct answer for the data is input to the input layer, the parameters of the function in the neural network model are adjusted (learned) so that the data output approaches the correct answer of the training data. By repeatedly learning the neural network model, the error between the data output based on the input data and the correct answer can be reduced. 【0049】 [Knowledge information] In this specification, knowledge information refers to information that includes at least a portion of the knowledge that is the output of the intermediate layer of a model, the output of the output layer, the information of the intermediate layer, or the information of the output layer. 【0050】 [Distillation of Knowledge] In this specification, knowledge distillation means a method of training another learning model, such as a student model, using the knowledge information of a teacher model, which is a learning model that has been trained based on training data. For example, knowledge distillation is a learning method in which the student model is trained using training data that has fewer types of training data than the number of types of training data used in the training of the teacher model, and the knowledge information created using the teacher model. 【0051】 [Primary model] In this specification, the primary model may be a primary bio-information distillation model. In this case, the primary bio-information distillation model is a student model created using knowledge information of a teacher model used for estimating emotions based on multiple types of bio-information, and used for estimating emotions based on multiple types of bio-information. For example, if a model used for estimating emotions based on multiple types of bio-information is defined as a teacher model, and a model created using knowledge information of the teacher model and used for estimating emotions based on multiple types of bio-information is defined as a student model, then the primary model in this specification may be either the teacher model or the student model described above. [Effects of the Invention] 【0052】 According to one embodiment of the present invention, it is possible to provide an emotion estimation device that enhances the subject's sense of satisfaction and improves the accuracy of emotion estimation while suppressing a decrease in the design flexibility of hardware resources. [Brief explanation of the drawing] 【0053】 [Figure 1] Figure 1 shows a schematic diagram illustrating the learning process by knowledge distillation in an emotion estimation device according to Embodiment 1 of the present invention, and the overall configuration of an emotion estimation device having a trained emotion estimation model. [Figure 2] Figure 2 shows the process diagrams for MAML-based learning and fine-tuning-based learning in the emotion estimation device. [Figure 3] Figure 3 is a schematic diagram showing the inner loop in MAML-based learning in an emotion estimation device. [Figure 4] Figure 4 is a schematic diagram illustrating the outer loop and fine-tuning-based learning in MAML-based learning for emotion estimation devices. [Figure 5] Figure 5 is a schematic diagram showing the configuration of an emotion estimation device according to Embodiment 2 of the present invention. [Figure 6] Figure 6 is a schematic diagram illustrating MAML-based learning and fine-tuning-based learning in an emotion estimation device according to Embodiment 1 of the present invention. [Modes for carrying out the invention] 【0054】 The following describes each embodiment with reference to the drawings. In each drawing, the same parts are denoted by the same reference numerals, and the description of those parts will not be repeated. Note that the dimensions of the components in each drawing do not faithfully represent the dimensions of the actual components or the dimensional ratios of each component. 【0055】 [Embodiment 1] <Overall configuration of the emotion estimation device> The emotion estimation device 1 according to Embodiment 1 of the present invention will be described using Figure 1. Figure 1 shows a schematic diagram illustrating the learning by knowledge distillation in the emotion estimation device 1 according to Embodiment 1 of the present invention, and the overall configuration of the emotion estimation device 1 having a trained emotion estimation model 30. 【0056】 As shown in Figure 1, the emotion estimation device 1 according to Embodiment 1 of the present invention estimates the emotion of subject U or subject T using an emotion estimation model 30 that has been trained based on the subject U's biometric information Bi (hereinafter simply referred to as "biometric information Bi") or the subject T's biometric information BiL (hereinafter simply referred to as "learning biometric information BiL"). In this embodiment, the emotion estimation device 1 creates secondary estimated emotion information Ei2 or tertiary estimated emotion information Ei3, which are emotion information relating to the subject U's emotion, from at least one type of biometric information Bi of subject U. 【0057】 In each of the following embodiments, the model used to estimate emotion based on multiple types of different biometric information Bi is defined as the first-order model 100. The first-order model 100 outputs first-order knowledge information Ki1 based on multiple types of biometric information Bi. Furthermore, the machine learning model created by learning using the first-order knowledge information Ki1 obtained based on the first-order model 100 is defined as the second-order biometric information distillation model 31. In addition, the model created by learning using the second-order knowledge information Ki2 obtained based on the second-order biometric information distillation model 31 is defined as the third-order biometric information distillation model 37. Thus, the model created by learning using the n-1-th order knowledge information obtained based on the n-1-th order biometric information distillation model is defined as the n-th order biometric information distillation model (where n is 2 or an integer greater than 2). n is the number representing the generation of the biometric information distillation model, counting from the first-order model, with the first-order model being the first generation. 【0058】 The emotion estimation device 1 includes a control unit 20 and a memory unit 40. The control unit 20 creates secondary estimated emotion information Ei2 based on biometric information Bi or secondary learning biometric information BiL2 input from an external source, and creates tertiary estimated emotion information Ei3 based on tertiary learning biometric information BiL3. 【0059】 The control unit 20 has an emotion estimation model 30 used to estimate the emotions of the subject U based on biometric information Bi. The control unit 20 is also configured to accept input of biometric information Bi, secondary learning biometric information BiL2, or tertiary learning biometric information BiL3 from an external source. 【0060】 The control unit 20 stores various programs and data for controlling the emotion estimation device 1, which includes the emotion estimation model 30 and the memory unit 40. The control unit 20 is configured to transmit biometric information Bi, secondary learning biometric information BiL2, or tertiary learning biometric information BiL3, which are input from an external source, to the memory unit 40. The control unit 20 is also configured to acquire various information, including biometric information Bi, secondary learning biometric information BiL2, or tertiary learning biometric information BiL3, and information related to emotions, which are stored in the memory unit 40. 【0061】 The emotion estimation model 30 is used to estimate the emotions of subject U based on biometric information Bi. The emotion estimation model 30 is a machine learning model trained on multiple training data sets, each combining multiple sets of secondary learning biometric information BiL2 or tertiary learning biometric information BiL3, and self-reported secondary learning self-emotion evaluation information EiL2 or tertiary learning self-emotion evaluation information EiL3 from multiple subjects T. The emotion estimation model 30 also includes a secondary biometric information distillation model 31 created using primary knowledge information Ki1 obtained based on the primary model 100, or a tertiary biometric information distillation model 37 created using secondary knowledge information Ki2 obtained based on the secondary biometric information distillation model 31. 【0062】 The secondary bioinformation distillation model 31 is used to create secondary estimated emotion information Ei2, which is information indicating the emotions of the subject U, based on bioinformation Bi. The secondary bioinformation distillation model 31 is configured to accept bioinformation Bi and secondary learning bioinformation BiL2 as inputs. The secondary bioinformation distillation model 31 is configured to output secondary estimated emotion information Ei2. 【0063】 The tertiary bioinformation distillation model 37 is used to create tertiary estimated emotion information Ei3, which is information indicating the emotions of the subject U, based on bioinformation Bi. The tertiary bioinformation distillation model 37 is configured to accept bioinformation Bi and tertiary learning bioinformation BiL3 as inputs. The tertiary bioinformation distillation model 37 is configured to output tertiary estimated emotion information Ei3. 【0064】 The control unit 20 is configured to output secondary estimated emotion information Ei2 created using the secondary bioinformation distillation model 31 of the emotion estimation model 30, or tertiary estimated emotion information Ei3 created using the tertiary bioinformation distillation model 37 of the emotion estimation model 30. The control unit 20 is configured to output the secondary estimated emotion information Ei2 or tertiary estimated emotion information Ei3 output by the emotion estimation model 30 to the storage unit 40 and to the outside. 【0065】 The memory unit 40 stores biological information Bi, biological information BiL2 for secondary learning, biological information BiL3 for tertiary learning, secondary estimated emotion information Ei2, and tertiary estimated emotion information Ei3, etc. The memory unit 40 has RAM, processor internal memory, hard disk, etc., for storing biological information Bi, biological information BiL2 for secondary learning, biological information BiL3 for tertiary learning, secondary estimated emotion information Ei2, and tertiary estimated emotion information Ei3, etc. The memory unit 40 is configured to output the stored biological information Bi, biological information BiL2 for secondary learning, biological information BiL3 for tertiary learning, secondary estimated emotion information Ei2, and tertiary estimated emotion information Ei3 to the control unit 20. 【0066】 In the emotion estimation device 1 configured in this way, the control unit 20 stores the biometric information Bi input from an external source in the storage unit 40. The control unit 20 also uses the emotion estimation model 30 to create secondary estimated emotion information Ei2 or tertiary estimated emotion information Ei3 of the subject U based on the biometric information Bi. The control unit 20 outputs the created secondary estimated emotion information Ei2 to the storage unit 40 or to an external source at intervals of time. 【0067】 <Creation of Primary Model 100> Next, the primary model 100 necessary for creating the emotion estimation model 30 will be described. The trained primary model 100 is a model that provides the emotion estimation model 30 with primary knowledge information Ki1 created based on multiple different types of biometric information Bi. In this embodiment, the primary model 100 is a training model for the secondary biometric information distillation model 31. 【0068】 The first-order model 100 is a machine learning model used to estimate emotions based on multiple different types of training biometric information BiL. The first-order model 100 is a model trained on training data consisting of a combination of first-order training self-emotion evaluation information EiL1, which is information about the emotions of subject T that is self-reported, and first-order training biometric information BiL1, which includes multiple different types of biometric information. The first-order training self-emotion evaluation information EiL1 is information related to subject T's emotions when subject T's first-order training biometric information BiL1 was measured, and is the correct emotion in the training data. In other words, the first-order training self-emotion evaluation information EiL1 is the emotion used to check the estimation accuracy (error) of the first-order estimated emotion information Ei1, which is estimated based on the input first-order training biometric information BiL1, during the training of the first-order model 100. 【0069】 The first-order model 100 is used to create the first-order estimated emotion information Ei1 based on the first-order learning self-emotion evaluation information EiL1. In the first-order model 100, the parameters of the first-order model are adjusted so that the error between the first-order estimated emotion information Ei1 and the first-order learning self-emotion evaluation information EiL1 is minimized. 【0070】 <Creation of a secondary bio-information distillation model 31> Next, the creation of the secondary bio-information distillation model 31 included in the emotion estimation model 30 will be described. The secondary bio-information distillation model 31 is created by learning based on primary knowledge information Ki1, secondary learning bio-information BiL2, and secondary learning self-emotion evaluation information EiL2. In this embodiment, primary estimated emotion information Ei1 is input to the secondary bio-information distillation model 31 as primary knowledge information Ki1. In this embodiment, the secondary bio-information distillation model 31 is a student model for the primary model 100. 【0071】 The emotion estimation device 1 is configured to output secondary estimated emotion information Ei2 when secondary learning biometric information BiL2 is input, using a secondary biometric information distillation model 31. The parameters of the secondary biometric information distillation model 31 are adjusted so that the sum of the error between the secondary estimated emotion information Ei2 and the secondary learning self-emotion evaluation information EiL2 contained in the training data, and the error between the secondary estimated emotion information Ei2 and the primary estimated emotion information Ei1, which is primary knowledge information Ki1, is as small as possible. The learning of the secondary biometric information distillation model 31 is performed based on the primary estimated emotion information Ei1 obtained using the primary model 100 based on the primary learning self-emotion evaluation information EiL1, the secondary estimated emotion information Ei2 obtained using the secondary biometric information distillation model 31 based on the secondary learning biometric information BiL2, and the secondary learning self-emotion evaluation information EiL2 of the training data. 【0072】 The secondary bioinformation distillation model 31 is used to estimate the emotions of subject U based on secondary learning bioinformation BiL2, which is included in the types of primary learning bioinformation BiL1 used when obtaining primary knowledge information Ki1 (when estimating emotions) using the primary model 100, and which has fewer types of primary learning bioinformation BiL1 than the number of types of primary learning bioinformation BiL1 required for emotion estimation in the primary model 100. 【0073】 The primary knowledge information Ki1 is the output of the primary model 100, obtained using the primary model 100 and based on multiple types of primary learning biometric information BiL1. In other words, the primary knowledge information Ki1 is the information (knowledge of the primary model 100) obtained by the primary model 100 through learning based on multiple types of primary learning biometric information BiL1 for estimating the emotions of subject U. 【0074】 The training of the secondary bioinformation distillation model 31 is performed by utilizing the primary knowledge information Ki1, thereby inheriting the primary estimated emotion information Ei1, which contains information (privileged information) not included in the secondary learning bioinformation BiL2 used to train the secondary bioinformation distillation model 31 (knowledge distillation). 【0075】 Therefore, the secondary bioinformation distillation model 31, which is trained using primary knowledge information Ki1 in addition to secondary learning bioinformation BiL2, can achieve higher emotion estimation accuracy than the model trained using only secondary learning bioinformation BiL2. Furthermore, even when the secondary bioinformation distillation model 31 is trained based on a smaller number of types of secondary learning bioinformation BiL2 than the number of types of primary learning bioinformation BiL1 used by the primary model 100 to obtain primary knowledge information Ki1, it can achieve an estimation accuracy close to that of the primary model 100. 【0076】 Furthermore, when the emotion estimation device 1 uses primary knowledge information Ki1 to train the secondary bioinformation distillation model 31, it can reduce the size of the model and the amount of bioinformation required for emotion estimation compared to when the secondary bioinformation distillation model 31 is trained without using primary knowledge information Ki1. Therefore, the emotion estimation device 1 can estimate the emotions of subject U with an estimation accuracy close to that of the primary model 100 by using primary knowledge information Ki1 while suppressing an increase in the processing load of the hardware. 【0077】 <Creation of a tertiary biosynthesis model> Next, the creation of the tertiary bioinformation distillation model 37 included in the emotion estimation model 30 will be described. The tertiary bioinformation distillation model 37 is created by learning based on tertiary knowledge information Ki2 obtained using a trained tertiary bioinformation distillation model 31 and based on multiple types of tertiary learning self-emotion evaluation information EiL2, and tertiary learning bioinformation BiL3. In this embodiment, the emotion estimation model 30 uses the tertiary estimated emotion information Ei2, created using a trained tertiary bioinformation distillation model 31 and based on tertiary learning bioinformation BiL2, as the tertiary knowledge information Ki2. 【0078】 The emotion estimation device 1 is configured to output tertiary estimated emotion information Ei3 when tertiary learning self-emotion evaluation information EiL3 is input, using a tertiary bioinformation distillation model 37. The parameters of the tertiary bioinformation distillation model 37 are adjusted so that the sum of the error between the tertiary estimated emotion information Ei3 and the tertiary learning self-emotion evaluation information EiL3 contained in the training data, and the error between the tertiary estimated emotion information Ei3 and the tertiary knowledge information Ki2, which is the tertiary estimated emotion information Ei2, is as small as possible. The learning of the tertiary bioinformation distillation model 37 is performed based on the tertiary learning bioinformation BiL2 obtained using a trained tertiary bioinformation distillation model 31, the tertiary estimated emotion information Ei3 obtained using the tertiary bioinformation distillation model 37 based on the tertiary learning bioinformation BiL3, and the tertiary learning self-emotion evaluation information EiL3 of the training data. 【0079】 The tertiary bioinformation distillation model 37 is used to estimate the emotions of subject U based on the tertiary learning bioinformation BiL3. The tertiary learning bioinformation BiL3 is included in the types of secondary learning bioinformation BiL2 used when obtaining secondary knowledge information Ki2 (when estimating emotions) using the tertiary bioinformation distillation model 31, and includes fewer types of bioinformation than the types of secondary learning bioinformation BiL2. 【0080】 The learning of the tertiary bioinformation distillation model 37, configured in this way, is performed based on the secondary estimated emotion information Ei2 obtained using the trained secondary bioinformation distillation model 31, the tertiary estimated emotion information Ei3 obtained using the tertiary bioinformation distillation model 37 based on the tertiary learning self-emotion evaluation information EiL3, and the tertiary learning self-emotion evaluation information EiL3. 【0081】 Secondary knowledge information Ki2 is the output of the secondary bioinformation distillation model 31, obtained using the secondary bioinformation distillation model 31 based on at least one type of secondary learning bioinformation BiL2. In other words, secondary knowledge information Ki2 is information (knowledge of the tertiary bioinformation distillation model 37) for estimating the emotions of subject U, obtained by the secondary bioinformation distillation model 31 through learning based on multiple different types of secondary learning bioinformation BiL2. 【0082】 The learning of the tertiary bioinformation distillation model 37 is performed by utilizing the secondary knowledge information Ki2, thereby inheriting the secondary estimated emotion information Ei2, which contains information (privileged information) that is not included in the tertiary learning self-emotion evaluation information EiL3 used for learning the tertiary bioinformation distillation model 37 (knowledge distillation). 【0083】 Therefore, the tertiary bioinformation distillation model 37, which is trained using secondary knowledge information Ki2 in addition to the tertiary self-emotion evaluation information EiL3, can achieve higher emotion estimation accuracy than the model trained using only the tertiary self-emotion evaluation information EiL3. Furthermore, even when the tertiary bioinformation distillation model 37 is trained based on a number of biometric information types less than or equal to the number of types of secondary learning biometric information BiL2 used by the secondary bioinformation distillation model 31 to obtain secondary knowledge information Ki2, it can achieve an estimation accuracy close to that of the secondary bioinformation distillation model 31. 【0084】 The emotion estimation device 1 configured in this way can estimate the emotions of subject U with an estimation accuracy close to that of the first-order model 100, while suppressing an increase in the processing load of the hardware, by utilizing the emotion estimation model 30 generated by knowledge distillation. For example, the emotion estimation device 1 can easily acquire biometric information Bi of subject U performing specific tasks or actions, such as driving an automobile, automobile, automobile, ship, airplane, various work machinery, drone, etc., walking, exercising such as running, or watching movies, paintings, or other content, using wearable sensors, etc., and estimate the emotions of subject U performing specific tasks or actions. 【0085】 The secondary bioinformation distillation model 31, created through knowledge distillation, can improve estimation accuracy even when convenience is enhanced by reducing the number of types of bioinformation Bi of the subject U used for emotion estimation. The secondary bioinformation distillation model 31 is created using primary knowledge information Ki1 obtained from a primary model created by complementaryly utilizing multiple types of bioinformation. This is thought to be because, when creating the secondary bioinformation model 31 that estimates emotions based on the bioinformation Bi of subject U, which has a small number of types, the primary knowledge information Ki1 functions to complement the bioinformation Bi of subject U, which has a small number of types. 【0086】 However, the secondary bioinformation distillation model is suitable for estimating emotions based on biometric information Bi obtained in an environment where secondary learning biometric information BiL2 was obtained from subject T, the subject whose secondary biometric information was used for training. Therefore, when using the secondary bioinformation distillation model 31 to estimate emotions based on biometric information Bi obtained in an environment different from the environment in which the secondary learning biometric information BiL2 used during training of the secondary bioinformation distillation model 31 was obtained, or based on biometric information Bix (see Figure 4) of a subject Ux different from subject T, the accuracy of emotion estimation may decrease. Therefore, in order to adapt the secondary bioinformation distillation model 31 to emotion estimation based on the biometric information Bix of subject Ux, the secondary bioinformation distillation model 31 is subjected to MAML-based learning and fine-tuning-based learning. 【0087】 <Individual Adaptation of Secondary Biomedical Information Distillation Model 31> Next, using Figures 2 to 4 and Figure 6, we will explain the adjustment of the secondary bio-information distillation model 31 included in the emotion estimation model 30 and the fitting of the tertiary bio-information distillation model 37 to the subject Ux. Figure 2 shows a process diagram for learning based on MAML and learning based on fine tuning in the emotion estimation device 1. Figure 3 is a schematic diagram showing the inner loop in learning based on MAML in the emotion estimation device 1. Figure 4 is a schematic diagram showing the outer loop and learning based on fine tuning in learning based on MAML in the emotion estimation device 1. Figure 6 is a schematic diagram showing learning based on MAML and learning based on fine tuning in the emotion estimation device 1. Note that the adjustment of the secondary bio-information distillation model 31 and the adjustment of the tertiary bio-information distillation model 37 are the same adjustment content, so in this embodiment, we will explain the adjustment of the secondary bio-information distillation model 31. 【0088】 As shown in Figure 2, the adaptation process of the secondary bioinformation distillation model 31 includes a learning process S1 based on MAML (Model-Agnostic Meta-Learning) and a learning process S2 based on fine-tuning. The learning process S1 based on MAML includes an information acquisition process S11 for MAML, an inner loop process S12 in MAML, and an outer loop process S13 in MAML. The learning process S2 based on fine-tuning includes an information acquisition process S21 for fine-tuning and a fine-tuning process S22. 【0089】 In the MAML-based learning process S1, the secondary bioinformation distillation model 31 is updated with initial parameters θM (see Figure 4) that can be easily adapted by MAML to the emotion estimation of any subject Ux. In the fine-tuning-based learning process S2, the secondary bioinformation distillation model 31 is updated by fine-tuning to parameters θx (see Figure 4) that are specialized for outputting emotion information of a specific subject Ux in a specific environment. In other words, the parameters of the secondary bioinformation distillation model 31 are adjusted to fit the subject Ux through learning based on MAML and fine-tuning. 【0090】 As shown in FIGS. 3 and 6, the emotion estimation device 1 performs MAML based on the biological information BiMa for meta-learning A of the subject Ta, the biological information BiMb for meta-learning B of the subject Tb, the biological information BiMc for meta-learning C of the subject Tc, the self-emotion evaluation information EiMa for meta-learning A of the subject Ta, the self-emotion evaluation information EiMb for meta-learning B of the subject Tb, and the self-emotion evaluation information EiMc for meta-learning C of the subject Tc. Further, the emotion estimation device 1 performs fine-tuning based on the biological information BiMx for meta-learning X of the target person Ux and the self-emotion evaluation information EiMx for meta-learning X. 【0091】 <Learning process based on MAML> As shown in FIGS. 3 and 6, in the MAML information acquisition step S11 (see FIG. 2) of the learning step S1 based on MAML, the emotion estimation device 1 uses the biological information BiMa for meta-learning A of the subject Ta, which is the same type of biological information as the biological information BiL2 used in the creation of the secondary biological information distillation model 31, and the self-emotion evaluation information EiMa for meta-learning A that evaluates the self-emotion information when the subject Ta measures the biological information BiMa for meta-learning A. Similarly, the emotion estimation device 1 acquires the biological information BiMb for meta-learning B and the self-emotion evaluation information EiMb for meta-learning B from the subject Tb, and acquires the biological information BiMc for meta-learning C and the self-emotion evaluation information EiMc for meta-learning C from the subject Tc. 【0092】 Furthermore, subjects Ta, Tb, and Tc are multiple providers different from subject T, who provided the biometric information BiL2 for secondary learning. In addition, the biometric information BiMa for meta-learning A, BiMb for meta-learning B, BiMc for meta-learning C, EiMa for meta-learning A, EiMb for meta-learning B, and EiMc for meta-learning C were acquired under different environments than the environment in which subject T's primary learning biometric information BiL1, used in the creation of the primary model 100, was acquired. Subject T's primary learning biometric information BiL1 is, for example, biometric information measured while viewing an image displayed on a screen indoors. Subject Ta's biometric information BiMa for meta-learning A is, for example, biometric information measured while driving a lean vehicle outdoors. 【0093】 In the inner loop process S12 (see Figure 2) of the MAML-based learning process S1, the emotion estimation device 1 learns a secondary bio-information distillation A model 31a, which estimates the emotion of subject Ta, one of the secondary bio-information distillation models 31, based on the acquired bio-information for meta-learning A BiMa and self-emotion evaluation information for meta-learning A EiMa. The emotion estimation device 1 calculates the error La between the estimated emotion information for meta-learning A Eia of subject Ta estimated using the secondary bio-information distillation A model 31a and the self-emotion evaluation information for meta-learning A EiMa. The emotion estimation device 1 updates (adjusts) the initial parameter θ of the secondary bio-information distillation A model 31a to parameter θa so as to minimize the error La using backpropagation. 【0094】 Similarly, in the inner loop process S12 of the MAML-based learning process S1, the emotion estimation device 1 learns a secondary bio-information distillation B model 31b that estimates the emotions of subject Tb based on the acquired bio-information BiMb for meta-learning B and self-emotion evaluation information EiMb for meta-learning B. The emotion estimation device 1 calculates the error Lb between the estimated emotion information Eib for meta-learning B and the self-emotion evaluation information EiMb for meta-learning B of subject Tb, estimated using the secondary bio-information distillation B model 31b. The emotion estimation device 1 updates the initial parameter θ of the secondary bio-information distillation B model 31b to parameter θb so as to minimize the error Lb using backpropagation. 【0095】 Similarly, in the inner loop process S12 of the MAML-based learning process S1, the emotion estimation device 1 learns a secondary bio-information distillation C model 31c that estimates the emotion of subject Tc based on the acquired bio-information BiMc for meta-learning C and self-emotion evaluation information EiMc for meta-learning C. The emotion estimation device 1 calculates the error Lc between the estimated emotion information Eic for meta-learning C of subject Tc estimated using the secondary bio-information distillation C model 31c and the self-emotion evaluation information EiMc for meta-learning C. The emotion estimation device 1 updates the initial parameter θ of the secondary bio-information distillation C model 31c to parameter θc so that the error Lc is minimized by backpropagation. 【0096】 As shown in Figures 4 and 6, in the outer loop process S13 (see Figure 2) of the MAML-based learning process S1, the emotion estimation device 1 updates the initial parameter θ of the second-order bioinformation distillation model 31 to the initial parameter θM by backpropagation so as to minimize the sum of the calculated errors La, Lb, and Lc. 【0097】 The second-order bio-distillation model 31, with its updated initial parameters θM based on the errors La, Lb, and Lc indicating the accuracy of emotion estimation for the second-order bio-distillation models A 31a, B 31b, and C 31c, can be easily adapted to estimating the emotions of subject U with various characteristics. In other words, the emotion estimation device 1 can create a subject Ux-specific second-order bio-distillation model X 31x using fewer types and amounts of bio-information than the learning bio-information BiL used for conventional machine learning, through MAML-based learning. 【0098】 <Learning process based on fine-tuning> As shown in Figures 4 and 6, in the fine-tuning information acquisition step S21 (see Figure 2) of the learning step S2 based on fine-tuning, the emotion estimation device 1 acquires biometric information BiMx for meta-learning X, which is the same type of biometric information as the secondary learning biometric information BiL2, and self-emotion evaluation information EiMx for meta-learning X, which is the subject Ux whose emotion is to be estimated. 【0099】 Furthermore, subject Ux is a different provider from subject T, who provided the biometric information BiL2 for secondary learning. In addition, the biometric information BiMx for meta-learning X and the self-emotion evaluation information EiMx for meta-learning X were acquired under different conditions than the environment in which the primary learning biometric information BiL1 used in the creation of the primary model 100 was acquired. 【0100】 In the fine-tuning step S22 (see Figure 2) of the learning step S2 based on fine-tuning, the emotion estimation device 1 learns a secondary bio-information distillation X model 31x, which is one of the secondary bio-information distillation models 31 and estimates the emotion of the subject Ux, based on the acquired bio-information BiMx for meta-learning X and self-emotion evaluation information EiMx for meta-learning X. The secondary bio-information distillation X model 31x has an initial parameter θM that was updated in the learning step S1 based on MAML. The emotion estimation device 1 calculates the error Lx between the estimated emotion information Eix for meta-learning X of the subject Ux estimated using the secondary bio-information distillation X model 31x and the self-emotion evaluation information EiMx for meta-learning X. The emotion estimation device 1 updates the initial parameter θM of the secondary bio-information distillation X model 31x to parameter θx so as to minimize the error Lx using backpropagation. 【0101】 The emotion estimation device 1 uses the secondary bioinformation distillation X model 31x to output secondary estimated emotion information Ei2 that better reflects the characteristics related to the subject Ux's bioinformation Bix and emotion, based on the subject Ux's bioinformation Bix. The bioinformation Bix is ​​included in the types of primary learning bioinformation BiL1 required for emotion estimation in the primary model 100, and includes fewer types of bioinformation than the number of types of primary learning bioinformation BiL1. 【0102】 The emotion estimation device 1 trains the secondary bioinformation distillation A model 31a based on the subject Ta's biometric information BiMa for meta-learning A and self-emotion evaluation information EiMa for meta-learning A. The emotion estimation device 1 also trains the secondary bioinformation distillation B model 31b based on the subject Tb's biometric information BiMb for meta-learning B and self-emotion evaluation information EiMb for meta-learning B. The emotion estimation device 1 also trains the secondary bioinformation distillation C model 31c based on the subject Tc's biometric information BiMc for meta-learning C and self-emotion evaluation information EiMc for meta-learning C. Based on the training results of each model, the emotion estimation device 1 updates the initial parameter θ of the secondary bioinformation distillation X model 31 to the initial parameter θM. Furthermore, the emotion estimation device 1 updates the initial parameter θM to the parameter θx through fine-tuning of the secondary bioinformation distillation X model 31x. Therefore, the emotion estimation device 1 can train the secondary bioinformation distillation X model 31x based on a smaller variety and amount of data than that used in conventional machine learning. 【0103】 The control unit 20 of the emotion estimation device 1 configured in this way acquires biometric information BiMx for meta-learning X and self-emotion evaluation information EiMx for meta-learning X in an environment different from the environment in which the primary learning biometric information BiL1 used in the creation of the primary model 100 was acquired. The biometric information BiMx for meta-learning X is the same type of biometric information as the secondary learning biometric information BiL2 input to the secondary biometric information distillation model 31. The self-emotion evaluation information EiMx for meta-learning X is self-emotion evaluation information in which the subject Ux evaluated their own emotions when the biometric information BiMx for meta-learning X was measured. Alternatively, the control unit 20 acquires biometric information BiMx for meta-learning X and self-emotion evaluation information EiMx for meta-learning X from a subject Ux different from subject T of the primary learning biometric information BiL1 used in the creation of the primary model 100. The biometric information BiMx for meta-learning X is the same type of biometric information as the secondary learning biometric information BiL2 input to the secondary biometric information distillation model 31. EiMx, the self-emotional assessment information for meta-learning X, is self-emotional assessment information obtained when the subject Ux evaluated their own emotions while the biometric information BiMx for meta-learning X was measured. 【0104】 Next, the control unit 20 uses a secondary bioinformation distillation X model 31x, which has been adjusted to reflect the characteristics of the different environment or the different subject Ux based on the acquired bioinformation BiMx for meta-learning X and self-emotion evaluation information EiMx for meta-learning X, to output secondary estimated emotion information Ei2 that better reflects the characteristics of the different environment or the different subject Ux, based on the bioinformation BiMx for meta-learning X of the subject Ux, which is included in the types of primary learning bioinformation BiL1 required for emotion estimation in the primary model 100 and has fewer types than the number of types of primary learning bioinformation BiL1 required for emotion estimation in the primary model 100. 【0105】 As described above, the biometric information BiMx for meta-learning X is of the same type as the biometric information BiL2 for secondary learning, and is biometric information Bix of a different subject Ux than the one obtained in the environment in which the biometric information BiL1 for primary learning was acquired. Furthermore, the self-emotion evaluation information EiMx for meta-learning X is self-emotion evaluation information of subject Ux in biometric information of the same type as the biometric information BiL2 for secondary learning. The emotion estimation device 1 adjusts the initial parameter θM of the secondary biometric information distillation model 31 based on the biometric information BiMx for meta-learning X and the self-emotion evaluation information EiMx for meta-learning X using MAML. As a result, since it is not necessary to increase the types of biometric information in order to improve the accuracy of emotion estimation, the reduction in the design degree of freedom of the hardware resources in the emotion estimation device 1 can be suppressed. 【0106】 Furthermore, the output secondary estimated emotion information Ei2 utilizes a secondary bio-information distillation X model 31x that has been adjusted to reflect the characteristics of the subject Ux based on the bio-information BiMx for meta-learning X and the self-emotion evaluation information EiMx for meta-learning X, thus better reflecting the characteristics of the subject Ux. In this way, the secondary estimated emotion information Ei2 is output using a secondary bio-information distillation X model 31x that takes into account the relationship between the bio-information BiMx for meta-learning X and the self-emotion evaluation information EiMx for meta-learning X, which is the cognitive emotion recognized by the subject Ux in the brain. As a result, the discrepancy between the secondary estimated emotion information Ei2 estimated based on the subject Ux's bio-information BiMx and the self-emotion evaluation information EiMx for meta-learning X is suppressed, increasing the subject Ux's sense of satisfaction and further improving the accuracy of emotion estimation. 【0107】 In other words, instead of increasing the types or amount of biometric information Bi data, by adjusting the parameters of the emotion estimation model 30 created by knowledge distillation using MAML and fine-tuning, it is possible to increase the sense of acceptance (understanding) of the subject Ux's secondary estimated emotion information Ei2 based on a small amount of biometric information Bi data for a given situation or subject Ux. As a result, the discrepancy between the secondary estimated emotion information Ei2 and the self-emotion evaluation information is suppressed, thereby improving the accuracy of the relative emotion estimation for the subject Ux. Therefore, in the emotion estimation device 1 that performs emotion estimation of subject Ux using the emotion estimation model 30, it is possible to realize a configuration that can further improve the accuracy of emotion estimation so that the subject Ux's sense of acceptance increases, while suppressing a decrease in the design flexibility of the hardware resources. 【0108】 [Embodiment 2] <Emotion estimation based on electrocardiogram signals and pulse waves> <Overall configuration of the emotion estimation device> The emotion estimation device 1 according to Embodiment 2 of the present invention will be described using Figure 5. Figure 5 is a schematic diagram showing the configuration of the emotion estimation device 1 according to Embodiment 2 of the present invention. In the following embodiments, specific explanations of points that are the same as those of the embodiments already described will be omitted, and the explanation will focus on the differences. 【0109】 As shown in Figure 5, the emotion estimation device 1 according to Embodiment 2 of the present invention estimates the emotions of a subject U based on biological information Bi. In this embodiment, the emotion estimation device 1 creates emotion information, which is information relating to the emotions of the subject U, from biological information Bi relating to the electrocardiogram signal and pulse wave of the subject U. 【0110】 The emotion estimation device 1 comprises a biometric information acquisition unit 10, a control unit 20, a storage unit 40, and an output unit 50. Of the emotion estimation device 1, the storage unit 40, the output unit 50, and the control unit 20 are built into a mobile terminal Pt such as a smartphone possessed by the subject U. 【0111】 In other words, the emotion estimation device 1 shares some of the components of the mobile terminal Pt. Specifically, the memory of the mobile terminal Pt functions as the storage unit 40 of the emotion estimation device 1. The liquid crystal monitor of the mobile terminal Pt functions as the output unit 50 of the emotion estimation device 1. The processor of the mobile terminal Pt functions as the control unit 20 of the emotion estimation device 1. 【0112】 The biological information acquisition unit 10 acquires (measures) biological information Bi or learning biological information BiL. The biological information acquisition unit 10 also outputs the acquired biological information Bi or learning biological information BiL to the control unit 20. The biological information acquisition unit 10 includes a sensor 11, such as a heart rate sensor (not shown), which acquires biological information Bi or learning biological information BiL, a control signal, and a sensor communication device 12 that transmits and receives the acquired biological information Bi or learning biological information BiL. 【0113】 The biometric information acquisition unit 10 uses the sensor 11 to acquire at least one of the following, such as biometric information Bi or learning biometric information BiL: electrocardiogram signal, heart rate, heart sounds, heart rate waveform, heart rate cycle, heart rate change, blood pressure, pulse wave, 3-axis acceleration, body surface temperature, electroencephalogram, respiratory rate, pupil state, electromyography, blood components, exhaled breath, exhaled breath volume, or exhaled breath components. 【0114】 In this embodiment, the biometric information acquisition unit 10 acquires biometric information Bi and learning biometric information BiL, for example, biometric information relating to the heart rate, specifically electrocardiogram signals and pulse waves. The biometric information acquisition unit 10 acquires the electrocardiogram signal and pulse wave of the subject U or subject T at unit time intervals using a wearable sensor, for example, a sensor 11 that can be easily attached by the subject U or subject T. The biometric information acquisition unit 10 also outputs the biometric information Bi, including the electrocardiogram signal and pulse wave of the subject U or subject T acquired by the sensor 11, to the control unit 20 using a sensor communication device 12. 【0115】 The control unit 20 controls the biometric information acquisition unit 10, the storage unit 40, and the output unit 50. The control unit 20 also creates estimated emotion information based on the biometric information Bi of the subject U input from the biometric information acquisition unit 10. 【0116】 The control unit 20 has an emotion estimation model 30 used to estimate the emotions of subject U or subject T based on the electrocardiogram signal and pulse wave of subject U or subject T. The control unit 20 also has a control unit communication device that sends and receives control signals, etc., to and from the biological information acquisition unit 10 to acquire the electrocardiogram signal and pulse wave of subject U or subject T. 【0117】 The control unit 20 stores various programs and data for controlling the biometric information acquisition unit 10, the memory unit 40, and the emotion estimation model 30. The control unit 20 is configured to send and receive control signals to and from the biometric information acquisition unit 10 via a control unit communication device. Similarly, the control unit 20 is configured to send and receive control signals to and from the memory unit 40. 【0118】 The control unit 20 is configured to transmit the biological information Bi or learning biological information BiL acquired by the biological information acquisition unit 10 to the storage unit 40. The control unit 20 is also configured to acquire various information, including the biological information Bi and emotional information of the subject U stored in the storage unit 40. The control unit 20 is also configured to output various information, including the biological information Bi and emotional information of the subject U, to the output unit 50. 【0119】 The emotion estimation device 1 configured in this way can estimate emotions without restricting at least one of the surrounding environment of the subject U, the behavior of the subject U, or the work of the subject U, by utilizing biometric information Bi that is less susceptible to the influence of the environment surrounding the subject U, the work of the subject U, or the actions of the subject U, biometric information Bi that can be easily obtained from the subject U by a wearable sensor, etc., or biometric information Bi that is less susceptible to the influence of the environment surrounding the subject U, the work of the subject U, or the actions of the subject U, and can be easily obtained from the subject U by a wearable sensor, etc. 【0120】 The emotion estimation device 1 can easily acquire electrocardiogram signals and pulse waves from subject U performing specific tasks or actions, such as driving a four-wheeled vehicle, two-wheeled vehicle, ship, airplane, various work machinery, drone, etc., walking, exercising (running, etc.), watching movies, paintings, content, etc., or performing specific tasks or actions, using wearable sensors, etc. By using biometric information Bi, which can be easily acquired from subject U, for emotion estimation, it is possible to improve the convenience of acquiring biometric information Bi when estimating emotions while improving the accuracy of emotion estimation. 【0121】 [Other embodiments] Furthermore, in each of the embodiments described above, the emotion estimation device 1 may indicate the emotions of subject U or subject T as secondary estimated emotion information Ei2 or tertiary estimated emotion information Ei3, which are information relating to the emotions of subject U or subject T, by a combination of the emotional arousal value and the emotional valence value. Alternatively, the emotion estimation device 1 may indicate the emotions of subject U or subject T as a ratio of each emotion. For example, the emotion estimation model may indicate the emotions of subject U or subject T as joy 0.85, enjoyment 0.12, sadness 0.02, anger 0.01, etc. 【0122】 Furthermore, in each of the embodiments described above, the emotion estimation model 30 may indicate the emotional valence value of subject U or subject T as a degree between positive and negative emotions, and the arousal value as a degree between a state of activated emotion and a state of deactivated emotion. Alternatively, the emotion estimation model 30 may classify the emotional valence value and arousal value of subject U or subject T into two categories: high level and low level. 【0123】 Furthermore, in each of the embodiments described above, ensemble learning is performed in the first-order model 100 to create first-order estimated sentiment information Ei1, which is the average of the estimated sentiment information created using the first internal model, the second internal model, and the third internal model. However, other types of ensemble learning may be performed on the first-order model. For example, in the first-order model, ensemble learning may be performed in which second estimated sentiment information is created using the second internal model based on the first estimated sentiment information created by the first internal model, and third estimated sentiment information is created using the third internal model based on the second estimated sentiment information. 【0124】 Furthermore, in each of the embodiments described above, the secondary bio-information distillation model 31 uses primary estimated emotion information Ei1 created using the primary model 100 as primary knowledge information Ki1. However, the secondary bio-information distillation model may be trained using the parameters of the intermediate layer of the trained primary model as primary knowledge information. Similarly, the tertiary bio-information distillation model may be trained using the parameters of the intermediate layer of the trained secondary bio-information distillation model as secondary knowledge information. 【0125】 Furthermore, in each of the embodiments described above, the primary model 100 is input with a multidimensional vector of learning biometric information BiL. However, the primary model may also be input with learning biometric information. In this case, the primary model is configured as a convolutional neural network model having convolutional layers. Also, the secondary biometric information distillation model 31 and the tertiary biometric information distillation model 37 are input with learning biometric information BiL. However, the secondary and tertiary biometric information distillation models may also be input with a multidimensional vector of learning biometric information. 【0126】 Furthermore, in each of the embodiments described above, the primary model 100 may have an encoder section and a first head section. Alternatively, the primary model 100 may be a model having at least one of the encoder section or the head section. 【0127】 Furthermore, in each of the embodiments described above, the first-order model 100, the second-order bio-information distillation model 31, and the tertiary bio-information distillation model 37 are composed of neural network models. However, the first-order model, the second-order bio-information distillation model, and the tertiary bio-information distillation model may be composed of machine learning models such as support vector machines (SVMs), decision trees, k-nearest neighbors, simple Bayesian methods, logistic regression, linear regression, nonlinear regression, and stepwise regression. 【0128】 In each of the embodiments described above, the emotion estimation device 1 is composed of a mobile terminal such as a smartphone held by the subject U and a wearable sensor. However, the emotion estimation device may be located on a server. The emotion estimation device estimates emotions based on biometric information transmitted from a mobile terminal such as a smartphone connected to the server. The emotion estimation device transmits the estimated emotion information to the mobile terminal. Furthermore, the emotion estimation device may be configured as part of an emotion estimation system in which multiple emotion estimation devices are connected to a server via the internet or the like. The emotion estimation system is composed of multiple emotion estimation devices connected via the internet or the like. Furthermore, the emotion estimation system may be configured to collect estimated emotion information of each subject U from multiple emotion estimation devices. 【0129】 In each of the embodiments described above, the emotion estimation model 30 includes a secondary bio-information distillation model 31 that has learned primary knowledge information Ki1 created using a trained primary model 100, or a tertiary bio-information distillation model 37 that has learned secondary knowledge information Ki2 created using a trained secondary bio-information distillation model 31. However, the emotion estimation model may also include a quaternary model that has learned tertiary knowledge information using information created using a trained tertiary bio-information distillation model. In other words, the emotion estimation model only needs to include a model that has learned information created using a trained model. Furthermore, the emotion estimation model may include models, programs, etc., other than the emotion estimation model. 【0130】 In each of the embodiments described above, the nth-order bioinformation distillation model is configured as a second-order bioinformation distillation model 31 and a third-order bioinformation distillation model 37. However, the nth-order bioinformation distillation model may also include a fourth-order bioinformation distillation model and a higher-order bioinformation distillation model than the fourth-order bioinformation distillation model. 【0131】 In each of the embodiments described above, subject Ux is a different provider from subject T, who is the provider of the secondary learning biometric information BiL2. Furthermore, the meta-learning biometric information BiMx and the meta-learning self-emotion evaluation information EiMx are acquired in an environment different from the environment in which the primary learning biometric information BiL1 used in the creation of the primary model 100 was acquired. However, the emotion estimation device may perform MAML-based learning and fine-tuning-based emotion estimation model training based on meta-learning biometric information measured by a subject different from the subject who is the provider of the nth-order learning biometric information, in the same environment as the subject. Alternatively, the emotion estimation device may perform MAML-based learning and fine-tuning-based emotion estimation model training based on meta-learning biometric information measured by the subject who is the provider of the nth-order learning biometric information, in an environment different from the environment in which the nth-order learning biometric information was measured. 【0132】 Although embodiments of the present invention have been described above, the embodiments described above are merely examples for carrying out the present invention. Therefore, the invention is not limited to the embodiments described above, and it is possible to carry out the invention by appropriately modifying the embodiments described above without departing from the spirit of the invention. [Explanation of symbols] 【0133】 1 Emotion estimation device 10. Biological Information Acquisition Unit 11 sensors 12. Sensor communication device 20 Control Unit 30. Emotion Estimation Models 31 Secondary Biomedical Information Distillation Model 31a Secondary Biomedical Information Distillation Model A 31b Secondary Biomedical Information Distillation Model B 31c Secondary Biomedical Information Distillation C Model 37. Third-order bio-information distillation model 40 Storage section 50 Output section Biological Information Biological Information for Learning Biological information for BiL1 primary training Biological information for secondary learning of BiL2 Biological information for BiL3 3rd learning EiL1 Primary Learning Self-Emotion Assessment Information Self-emotion assessment information for EiL2 secondary learning. Self-emotion assessment information for EiL3 3rd stage learning Ei1 Primary estimated sentiment information Ei2 Secondary estimated emotion information Ei3 3rd-order estimated emotion information Ki1 primary knowledge information Ki2 secondary knowledge information Eia A-Model Estimated Sentiment Rating Information Eib B-Model Estimated Sentiment Rating Information EIC C-Model Estimated Sentiment Rating Information Eix X Model Estimated Sentiment Rating Information Self-emotion assessment information for EiMa meta-learning A Self-emotion assessment information for EiMb meta-learning B EiMc Meta-learning C Self-Sentiment Assessment Information Self-emotion assessment information for EiMx meta-learning X Biometric information for BiMa meta-learning A Biometric information for BiMb meta-learning B BiMc Meta-learning for C Biometric Information Biometric information for BiMx meta-learning X 100 Primary Model La, Lb, Lc error Subjects T, Ta, Tb, Tc U, Ux target audience

Claims

[Claim 1] A memory unit that stores biological information, A control unit having an emotion estimation model used to estimate emotions based on the aforementioned biometric information, acquiring the subject's biometric information and storing it in the memory unit, using the emotion estimation model to estimate the subject's emotions based on the subject's biometric information stored in the memory unit, and outputting the estimated emotion information as emotion information, An emotion estimation device having, If we define a model used to estimate emotions based on multiple types of the aforementioned bio-information as a first-order model, and define a second-order bio-information distillation model created using the first-order knowledge information obtained by the first-order model, or a third-order bio-information distillation model created using the second-order knowledge information obtained by the second-order bio-information distillation model, and so on, as an n-th-order bio-information distillation model (where n is 2 or an integer greater than 2), then, The aforementioned emotion estimation model is The nth-order bioinformation distillation model is used to estimate the emotions of the subject, based on the number of types of bioinformation that are included in the types of bioinformation necessary for estimating emotions in the first-order model and which is fewer than the number of types of bioinformation necessary for estimating emotions in the first-order model. The control unit, An emotion estimation device that uses the emotion estimation model to estimate the emotions of a subject based on the subject's biometric information which is included in the types of biometric information necessary for estimating emotions in the first-order model and is fewer in number than the number of types of biometric information necessary for estimating emotions in the first-order model, and outputs the estimated emotion information as estimated emotion information, The control unit, Biological information of the same type as the biological information input to the nth-order biological information distillation model, measured under an environment different from the environment in which the biological information used to create the first-order model was acquired, and self-emotion evaluation information in which the provider of the biological information evaluates their own emotions at the time the biological information was measured. or Obtain biometric information of the same type as the biometric information input to the nth-order biometric information distillation model, measured from a provider different from the provider of biometric information used to create the n-1th-order biometric information distillation model, and obtain self-emotion evaluation information in which the provider of the biometric information evaluates their own emotions at the time the biometric information was measured. Based on the acquired biometric information and self-emotional assessment information, the nth-order biometric information distillation model, which has been adjusted to reflect the characteristics of the different environment or the different provider, is used to estimate emotions that better reflect the characteristics of the different environment or the different subject, based on the subject's biometric information, which is included in the types of biometric information necessary for estimating emotions in the first-order model and is fewer in number than the number of types of biometric information necessary for estimating emotions in the first-order model. The estimated emotion information is then output as estimated emotion information. Emotion estimation device. [Claim 2] In the emotion estimation device according to claim 1, The aforementioned n-th order bioinformation distillation model is, Using the initial parameters of the nth-order bioinformation distillation model, which is adjusted based on the bioinformation and self-emotional assessment information obtained under the aforementioned different environments or from the aforementioned different subjects, the model is adjusted to reflect the characteristics of the aforementioned different environments or the aforementioned different subjects. Emotion estimation device.

Images