Learning device and learning method

Abstract

A learning device according to one aspect of the present disclosure comprises: a division unit which divides time-series data contained in supplied training data into segments for each of predetermined time intervals; an estimation unit which, on the basis of the segments and an estimator for which a to-be-learned parameter is set, estimates, for each segment, a parameter of a probability distribution with which correct output information contained in the training data conforms; and a learning unit which, on the basis of the parameters of the probability distributions for each segment and the correct output information, learns the to-be-learned parameter in such a way as to minimise a loss function in the form of a likelihood function relating to a probability distribution represented by the product of the probability distributions for each segment.

Description

Learning device and learning method

[0001] The present disclosure relates to a learning device and a learning method.

[0002] There is a known regression problem that estimates real-valued output information from time-series information. For example, there is a known task that estimates a score (e.g., a score representing psychological safety) for the entire conversation based on time-series information of video and audio of two speakers in a conversation. In such a task, the time-series information is very high-dimensional, so a method is used in which the time-series information is divided into segments of a certain time interval and the estimation results of each segment are aggregated (e.g., Non-Patent Documents 1 and 2, etc.).

[0003] David Curto, Albert Clapes, Javier Selva, Sorina Smeureanu, Julio Junior, CS Jacques, David Gallardo-Pujol, Georgina Guilera, David Leiva, Thomas B Moeslund, et al., "Dyadformer: A multi-modal transformer for long-range modeling of dyadic interactions," in Proceedings of the IEEE / CVF international conference on computer vision, 2021, pp. 2177-2188.Cristina Palmero, Javier Selva, Sorina Smeureanu, Julio Junior, CS Jacques, Albert Clapes, Alexa Mosegui, Zejian Zhang, David Gallardo, Georgina Guilera, et al., "Context-aware personality inference in dyadic scenarios: Introducing the udiva dataset," in Proceedings of the IEEE / CVF Winter Conference on Applications of Computer Vision, 2021, pp. 1-12.

[0004] However, since the information that serves as a clue for estimating the output information is not necessarily contained uniformly in each segment, the estimation accuracy of the conventional method may be reduced.

[0005] The present disclosure has been made in consideration of the above points, and aims to accurately estimate output information corresponding to time-series information.

[0006] A learning device according to one aspect of the present disclosure includes a dividing unit that divides time series data included in given learning data into segments for each predetermined time interval; an estimator to which learning parameters are set; an estimation unit that estimates, for each segment, parameters of a probability distribution to which correct answer output information included in the learning data follows, based on the segments; and a learning unit that learns the learning parameters based on the parameters of the probability distribution for each segment and the correct answer output information, using a likelihood function related to a probability distribution expressed as the product of the probability distributions for each segment as a loss function, so as to minimize the loss function.

[0007] For time-series information, output information corresponding to the time-series information can be estimated with high accuracy.

[0008] FIG. 1 is a diagram showing an example of model learning in a conventional method. FIG. 2 is a diagram showing an example of inference in a conventional method. FIG. 3 is a diagram showing an example of model learning in proposed method 1. FIG. 4 is a diagram showing an example of inference in proposed method 1. FIG. 5 is a diagram showing an example of model learning in proposed method 2. FIG. 6 is a diagram showing an example of the hardware configuration of an estimation device according to this embodiment. FIG. 7 is a diagram showing an example of the functional configuration of an estimation device according to this embodiment during model learning. FIG. 8 is a flowchart (Example 1) showing an example of model learning processing according to this embodiment. FIG. 9 is a flowchart (Example 2) showing an example of the model learning processing according to this embodiment. FIG. 10 is a diagram showing an example of the functional configuration of an estimation device according to this embodiment during inference. FIG. 11 is a flowchart showing an example of inference processing according to this embodiment.

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

[0010] <Conventional Method> First, a conventional method will be described. Hereinafter, we consider a task of estimating a real score corresponding to time-series data, which is time-series information, as output information (this is also called a "time-series estimation problem"). An example of such a task is a task of estimating a score representing the psychological safety of an entire conversation based on time-series data of video and audio from one session between two speakers (e.g., 10-minute time-series data).

[0011] In tasks such as those described above, time series data is very high-dimensional information, making it difficult to process all of the time series data at once due to memory and other constraints. For this reason, conventional methods divide time series data into fixed-time segments, input each segment individually into an estimator, and then aggregate multiple estimation results (e.g., Non-Patent Documents 1 and 2). During model training, a single correct score is assigned to the time series data as the correct score for the estimation result of each segment, and the estimator parameters are trained to minimize the error between the estimation result of each segment and the correct score (Figure 1). Meanwhile, during inference, the time series data (test data) to be estimated is similarly divided into segments, each segment is input into a trained estimator, and the scores obtained as the estimation results for each segment are averaged to obtain an estimation result for the entire time series (Figure 2). The correct score may also be called, for example, a "correct label" or the like.

[0012] The conventional method will be described in more detail with reference to FIGS. 1 and 2. ｎ Let s be the time series data of session n. ｎ is the correct score for session n. ｎ and correct score s ｎ On the other hand, at the time of inference, a set of training data represented by the pair of time series data X ｎ is given as test data. In this case, the time series estimation problem is solved by ｎ Correct score from ｎ The purpose is to accurately estimate the time series. That is, the time series estimation problem is expressed by the following equation (1).

[0013] Here, s ｎ The symbol with a hat "^" directly above it represents the estimated score. Also, f(·) is the regression function defined by the model, and Θ is its learning target parameter. Hereinafter, in the text of this specification, the estimated score will be referred to as "^s ｎ " will be expressed as ".

[0014] Time series data X ｎ It may be difficult to process the entire data simultaneously due to limitations in calculation memory, etc. For this reason, in the conventional method, ｎ K short segments X ｎ，ｋ The output is estimated for each segment by the following equations (2) and (3): ｎ，ｋ is being calculated.

[0015] where f ｓｅｇｍｅｎｔ (·) is a segment-wise estimator (e.g., a function such as a transformer), θ ｓ is the parameter to be learned.

[0016] In this case, the estimator f ｓｅｇｍｅｎｔ (・) is the correct score s ｎ is the correct score for each segment s ｎ，ｋ The learning parameter θ ｓ is learned. That is, the estimator f ｓｅｇｍｅｎｔ (・) learning parameter Θ=θ ｓ is trained to minimize the loss function L shown in the following equation (4).

[0017] Here, l(·,·) is a loss function for point estimation, and for example, mean squared error (MSE) is used.

[0018] During inference, the time series data X ｎ Each segment X ｎ，ｋ The estimated score ^s ｎ，ｋ The estimation result for the entire time series is obtained by averaging the following. ｎ Calculate.

[0019] As a result, the time series data X ｎ Estimated score ^s for ｎ is obtained.

[0020] <Issues with conventional methods> In conventional methods, time series data X ｎCorrect score for ｎ The correct score for each segment is s ｎ，ｋ The estimated score for each segment, s, can be used as ｎ，ｋ and correct score s ｎ，ｋ The learning parameter Θ=θ is set to minimize the error between ｓ However, the output information (i.e., the correct score s ｎ Since the information that serves as a clue for estimating the psychological safety level is not necessarily uniformly contained in each segment, estimation accuracy may be reduced. For example, a segment that only contains a greeting during a conversation may be included in both high and low psychological safety conversations, and such a segment is likely to lack information about the psychological safety perceived from the conversation. As a result, during model training, the correlation between the segment content and the correct score may be low, which may cause estimation accuracy to deteriorate. Furthermore, during inference, estimated scores, which may have different reliability depending on the segment, are averaged with the same weight to obtain an estimation result for the entire time series, which may also cause estimation accuracy to deteriorate.

[0021] <Proposed Methods> Below, proposed methods 1 and 2 are proposed as methods for estimating the probability distribution of scores for each segment, rather than performing point estimation of the scores for each segment. In these proposed methods, during model training, a high-entropy probability distribution is estimated for segments that do not contain information that can serve as clues for estimating output information, thereby mitigating the impact of correct scores on model training. Also, during inference, a high-entropy probability distribution is estimated for segments that do not contain information that can serve as clues for estimating output information, thereby mitigating the impact of estimation results for segments on estimation results for time series units.

[0022] Proposed method 1: As shown in the following equation (6), the estimator f ｓｅｇｍｅｎｔ is μ ｎ，ｋ and log σ ｎ，ｋIt is also assumed that the correct score for the entire time series follows the product of Gaussian distributions estimated for each segment, as shown in the following equation (7).

[0023] Here, P(s ｎ |X ｎ , θ ｓ ) is the correct score s ｎ Generation probability of time series data X ｎ and the learning parameter θ ｓ When a correct score s is given ｎ Z = ∫Π ｋ N (s ｎ ; μ ｎ，ｋ , σ ｎ，ｋ ２ )ds ｎ is a regularization term, which is assumed to be a constant during model training to avoid algorithmic complexity. ｎ，ｋ and σ ｎ，ｋ ２ are the mean and variance of the Gaussian distribution, respectively. ｓｅｇｍｅｎｔ is μ ｎ，ｋ and σ ｎ，ｋ Alternatively, μ ｎ，ｋ and σ ｎ，ｋ ２ In the following, (μ ｎ，ｋ , σ ｎ，ｋ ２ ), (μ ｎ，ｋ , σ ｎ，ｋ ), (μ ｎ，ｋ , log σ ｎ，ｋ ) are called "probability distribution parameters."

[0024] However, the Gaussian distribution is just an example, and a probability distribution other than the Gaussian distribution may be assumed. ｎ may be assumed to follow a product of some predetermined probability distributions estimated for each segment.

[0025] <<Model Learning>> Model learning in proposed method 1 will be described with reference to FIG. 3. ｎ Let s be the time series data of session n.ｎ is the correct score for session n. ｎ and correct score s ｎ In this case, as in the conventional method, the time series data X ｎ K short segments X ｎ，ｋ Then, for each segment, an estimator f ｓｅｇｍｅｎｔ This allows us to calculate μ for each segment. ｎ，ｋ and log σ ｎ，ｋ is estimated.

[0026] Then, the estimator f ｓｅｇｍｅｎｔ Learning parameter Θ=θ ｓ is trained to minimize the loss function L shown in the following equation (8).

[0027] Here, P(s ｎ ; μ ｎ，ｋ , σ ｎ，ｋ ２ ) is the kth segment X ｎ，ｋ The loss function L shown in the above equation (8) is a probability density function of a Gaussian distribution corresponding to the mean μ ｎ，ｋ , variance σ ｎ，ｋ ２ It should be noted that the symbol "c" added directly above "=" in the above equation (8) indicates that both sides are equal except for the constant term.

[0028] As shown in equation (8) above, the estimation error is ｓｅｇｍｅｎｔ The variance σ estimated by ｎ，ｋ ２ Therefore, the variance σ ｎ，ｋ ２ For segments estimated to have large σ, the impact of the error on model training is mitigated.

[0029] <<Inference>> The inference of the proposed method 1 will be described with reference to FIG. 4. During inference, time series data X ｎ is given as test data. Similarly, during inference, the estimation target data X ｎ K short segments Xｎ，ｋ Then, for each segment, an estimator f ｓｅｇｍｅｎｔ This allows us to calculate μ for each segment. ｎ，ｋ and log σ ｎ，ｋ is estimated.

[0030] Then, the score with the highest generation probability is calculated as the estimated score ^s using the following equation (9): ｎ Output as

[0031] As shown in equation (9) above, the estimation result for the entire time series is calculated by the variance σ ｎ，ｋ ２ This means that the influence of the estimation results of segments that do not contain information that can be used as clues when estimating the output information is reduced.

[0032] Proposed Method 2: Correct scores often contain label noise. Label noise is caused by annotator characteristics, etc., and is difficult to estimate from segments. Hereinafter, correct scores that may contain label noise (i.e., correct scores given together with time-series data as training data) will be referred to as "observed scores." Proposed Method 1 estimates the probability distribution of scores from observed scores that may contain label noise, which may result in a deterioration in estimation accuracy. Therefore, Proposed Method 2 describes a method for estimating the probability distribution while also taking label noise into account.

[0033] Below is the time series data X ｎ For , the observation score is as follows:

[0034] On the other hand, the true unobserved score is s ｎ In the following text of this specification, the observation score will be referred to as " ～ s ｎ " will be expressed as ".

[0035] As in the proposed method 1, the proposed method 2 also uses the estimator f ｓｅｇｍｅｎｔ is the probability distribution parameter (μ ｎ，ｋ , log σｎ，ｋ ) and outputs the true correct score s ｎ is assumed to follow a product of Gaussian distributions estimated for each segment.

[0036] Note that Z = ∫Π ｋ N (s ｎ ; μ ｎ，ｋ , σ ｎ，ｋ ２ )ds ｎ is a regularization term, which is assumed to be a constant during model training, as in proposed method 1. Also, as in proposed method 1, the estimator f ｓｅｇｍｅｎｔ is the probability distribution parameter (μ ｎ，ｋ , σ ｎ，ｋ ), or the probability distribution parameter (μ ｎ，ｋ , σ ｎ，ｋ ２ ) may be output.

[0037] The distribution of label noise is modeled by the following equation (12): ～ s ｎ is the true correct score s ｎ Given a true correct score s ｎ We assume that the probability distribution depends only on the difference between

[0038] Here, P δ (·) represents the distribution of label noise. In the above formula (12), the distribution of label noise P δ As an example of (·), a Laplace distribution with a parameter λ is assumed. However, the Laplace distribution is only an example of the distribution of label noise, and a probability distribution other than the Laplace distribution may be assumed.

[0039] <<Model Learning>> Model learning of the proposed method 2 will be described with reference to FIG. 5. During model learning, time series data X ｎ and observation score ～ s ｎ A set of training data (training data set) is given, which is represented by the pair s and s. In this case, in model training, the true correct score s ｎ and the estimator f ｓｅｇｍｅｎｔ The learning parameter θ ｓFor simplicity, let us assume that X = X ｎ , ～ s = ～ s ｎ We will express this as follows.

[0040] Specifically, the training data set {(X, ～ s)}, maximize the posterior probability of {s = {s ｎ}, θ ｓ} is estimated.

[0041] This is equivalent to minimizing the loss function L shown in the following equation (13).

[0042] Here, e ｎ = ～ s ｎ -s ｎ As shown in the above formula (13), the loss function L is ｎ，ｋ +e ｎ ), variance σ ｎ，ｋ ２ It can be seen that this reduces to a Gaussian negative log-likelihood function of ∑ ∑ i = ...

[0043] Based on the loss function L shown in the above equation (13), e = {e ｎ} and the learning parameter θ ｓ and are updated by the stochastic gradient descent method. ｎ The update is performed by dividing the session number n in the training dataset by the label noise e ｎ The learning target parameter θ of the embed function shown in the following equation (14) is estimated. ｅｍｂｅｄ This is realized as an update of

[0044] where Embed(·) is the embed function, θ ｅｍｂｅｄ is the parameter to be learned.

[0045] From the above, the learning parameter Θ = {θ ｓ , θ ｅｍｂｅｄ} is trained to minimize the loss function L shown in equation (13) above.

[0046] <<During Inference>> The process during inference is the same as in Proposed Method 1, and therefore a description thereof will be omitted.

[0047] The following describes an estimation device 10 that realizes the model learning and inference of the above-described proposed method 1 or 2. Note that, although the following describes a case where the same estimation device 10 performs model learning and inference, model learning and inference may be realized by different devices. Furthermore, a device that realizes model learning may be called a "model learning device" or a "learning device," etc.

[0048] <Example of Hardware Configuration of Estimation Apparatus 10> An example of the hardware configuration of the estimation apparatus 10 according to this embodiment will be described with reference to Fig. 6. Fig. 6 is a diagram illustrating an example of the hardware configuration of the estimation apparatus 10 according to this embodiment.

[0049] 6 , the estimation device 10 according to this embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a random access memory (RAM) 105, a read only memory (ROM) 106, an auxiliary storage device 107, and a processor 108. These pieces of hardware are connected to each other via a bus 109 so as to be able to communicate with each other.

[0050] The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the estimation device 10 does not necessarily have to include at least one of the input device 101 and the display device 102, for example.

[0051] The external I / F 103 is an interface with an external device such as a recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

[0052] The communication I / F 104 is an interface for connecting to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory. The processor 108 is one of various arithmetic devices such as a central processing unit (CPU) or a graphics processing unit (GPU).

[0053] 6 is an example, and the hardware configuration of the estimation device 10 is not limited to this. The estimation device 10 may have, for example, multiple auxiliary storage devices 107 or multiple processors 108, may not have some of the hardware shown in the figure, or may have various hardware other than the hardware shown in the figure.

[0054] <Example of Functional Configuration of Estimation Device 10 During Model Learning> An example of the functional configuration of the estimation device 10 according to this embodiment during model learning will be described with reference to Fig. 7. Fig. 7 is a diagram illustrating an example of the functional configuration of the estimation device 10 according to this embodiment during model learning.

[0055] 7 , the estimation device 10 during model training includes a training dataset acquisition unit 201, a segment division unit 202, an estimation unit 203, and a model training unit 204. These units are realized, for example, by a process in which one or more programs installed in the estimation device 10 are executed by the processor 108 or the like. The estimation device 10 during model training also includes a training dataset storage unit 301 and a model parameter storage unit 302. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. However, at least one of the training dataset storage unit 301 and the model parameter storage unit 302 may be realized by a storage area of ​​a storage device (e.g., a storage device included in a database server) communicatively connected to the estimation device 10.

[0056] The learning dataset acquisition unit 201 acquires a learning dataset from the learning dataset storage unit 301. Hereinafter, the learning dataset is referred to as D. In the case of proposed method 1, D={(X ｎ , s ｎ ) |n=1, ..., |D|}. X ｎ is the time series data of session n, s ｎ is the correct score for session n. On the other hand, in the case of proposed method 2, D = {(X ｎ , ～ s ｎ ) |n=1, ..., |D|}. X ｎ is the time series data of session n, ～ s ｎ is the observed score for session n.

[0057] The segment division unit 202 divides each time series data X ｎ K segments X ｎ，ｋ (k=1, . . . , K).

[0058] The estimation unit 203 estimates the learning parameter θ included in the learning parameter Θ stored in the model parameter storage unit 302. ｓ and the estimator f ｓｅｇｍｅｎｔ and segment X ｎ，ｋ to probability distribution parameters (e.g., μ ｎ，ｋ , log σ ｎ，ｋ ), (μｎ，ｋ , σ ｎ，ｋ ), (μ ｎ，ｋ , σ ｎ，ｋ ２ ) etc.) are estimated.

[0059] The model learning unit 204 calculates the probability distribution parameters estimated by the estimation unit 203 and the correct score s ｎ or observation score ～ s ｎ The learning target parameters Θ stored in the model parameter storage unit 302 are learned so as to minimize the loss function L by using at least the above formula (8). In the case of proposed method 1, the loss function L shown in the above formula (13) is used, and in the case of proposed method 2, the loss function L shown in the above formula (14) is used.

[0060] The training data set storage unit 301 stores the training data set D given to it.

[0061] The model parameter storage unit 302 stores the learning target parameter Θ. In the case of proposed method 1, Θ=θ ｓ , In the case of proposed method 2, Θ = {θ ｓ , θ ｅｍｂｅｄ}.

[0062] <Model Learning Process> Example 1> The model learning process when model learning is performed by the proposed method 1 will be described with reference to Fig. 8. Fig. 8 is a flowchart (Example 1) showing an example of the model learning process according to this embodiment.

[0063] The learning dataset acquisition unit 201 acquires a learning dataset D={(X ｎ , s ｎ ) |n=1, . . . , |D|} is acquired (step S101).

[0064] The segment division unit 202 divides each time series data X ｎ K segments X ｎ，ｋ (k=1, . . . , K) (step S102).

[0065] The estimation unit 203 estimates the learning target parameter Θ=θ stored in the model parameter storage unit 302. ｓ and the estimator f ｓｅｇｍｅｎｔand segment X ｎ，ｋ to probability distribution parameters (e.g., μ ｎ，ｋ , log σ ｎ，ｋ ), (μ ｎ，ｋ , σ ｎ，ｋ ), (μ ｎ，ｋ , σ ｎ，ｋ ２ ) etc.) are estimated (step S103).

[0066] The model learning unit 204 calculates the probability distribution parameters estimated in step S103 and the correct score s ｎ and the learning target parameters Θ=θ stored in the model parameter storage unit 302 are calculated so as to minimize the loss function L shown in the above equation (8). ｓ (Step S104). Note that any known optimization method (e.g., stochastic gradient descent) may be used to update the learning parameter Θ.

[0067] Example 2 A model learning process when model learning is performed by proposed method 2 will be described with reference to Fig. 9. Fig. 9 is a flowchart (Example 2) showing an example of the model learning process according to this embodiment.

[0068] The learning dataset acquisition unit 201 acquires a learning dataset D={(X ｎ , ～ s ｎ ) |n=1, . . . , |D|} is acquired (step S201).

[0069] The segment division unit 202 divides each time series data X ｎ K segments X ｎ，ｋ (k=1, . . . , K) (step S202).

[0070] The estimation unit 203 estimates the learning parameter θ included in the learning parameter Θ stored in the model parameter storage unit 302. ｓ and the estimator f ｓｅｇｍｅｎｔ and segment X ｎ，ｋ The probability distribution parameters are estimated from (step S203).

[0071] The model learning unit 204 uses the probability distribution parameters estimated in step S203 and the observation score ～ s ｎ , the session number n, and the embed function Embed(·) are used to calculate the learning target parameters Θ={θ ｓ , θ ｅｍｂｅｄ} is updated (learned) (step S204). Note that any known optimization method (e.g., stochastic gradient descent) may be used to update the learning parameter Θ.

[0072] <Example of Functional Configuration of Estimation Device 10 During Inference> An example of the functional configuration of the estimation device 10 according to this embodiment during inference will be described with reference to Fig. 10. Fig. 10 is a diagram illustrating an example of the functional configuration of the estimation device 10 according to this embodiment during inference.

[0073] 10 , the estimation device 10 at the time of inference includes a test data acquisition unit 205, a segment division unit 206, an estimation unit 207, and an output unit 208. These units are realized, for example, by a process in which one or more programs installed in the estimation device 10 are executed by the processor 108 or the like. The estimation device 10 at the time of inference also includes a test data storage unit 303 and a trained model parameter storage unit 304. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. However, at least one of the test data storage unit 303 and the trained model parameter storage unit 304 may be realized by a storage area of ​​a storage device (e.g., a storage device included in a database server) communicatively connected to the estimation device 10.

[0074] The test data acquisition unit 205 acquires the test data X from the test data storage unit 303. ｎ The test data X ｎ is the time series data for which the score is to be estimated.

[0075] The segment division unit 206 divides the test data X ｎ K segments X ｎ，ｋ (k=1, . . . , K).

[0076] The estimation unit 207 estimates the learned parameters θ ｓ and the estimator f ｓｅｇｍｅｎｔ and the estimated score ^s ｎ That is, the estimation unit 207 calculates the estimated score ^s using the above equation (9). ｎ Calculate.

[0077] The output unit 208 outputs the estimated score ^s calculated by the estimation unit 207. ｎ The output destination is not limited to a specific output destination and can be any output destination, but may be, for example, a storage area such as the auxiliary storage device 107, the display device 102 such as a display, or another device or equipment connected so as to be able to communicate.

[0078] The test data storage unit 303 stores the given test data X ｎ Remember.

[0079] The learned model parameter storage unit 304 stores the learned parameters θ ｓ The learned parameters θ ｓ is the trained estimator f ｓｅｇｍｅｎｔ parameter θ ｓ This is what I mean.

[0080] <Inference Processing> The inference processing when performing inference using proposed method 1 or 2 will be described with reference to Fig. 11. Fig. 11 is a flowchart showing an example of the inference processing according to this embodiment.

[0081] The test data acquisition unit 205 acquires the test data X from the test data storage unit 303. ｎ is acquired (step S301).

[0082] The segment division unit 206 divides the test data X ｎ K segments X ｎ，ｋ (k=1, . . . , K) (step S302).

[0083] The estimation unit 207 estimates the learned parameters θ ｓ and the estimator f ｓｅｇｍｅｎｔ Using the above equation (9), the estimated score ^s ｎ is calculated (step S303).

[0084] The output unit 208 outputs the estimated score ^s calculated in step S303. ｎ is output to a predetermined output destination (step S304).

[0085] As described above, the estimation device 10 according to the present embodiment estimates output information for each segment from time-series information and then aggregates the output information to estimate a probability distribution of the output information for each segment. This makes it possible to mitigate the impact on model learning and inference of segments that do not contain information that serves as a clue for estimating the output information, and is expected to improve estimation accuracy.

[0086] Furthermore, the estimation device 10 according to this embodiment is expected to improve estimation accuracy by simultaneously estimating the true correct score and the probability distribution of the output information, even in cases where label noise may be included in the correct score.

[0087] The present invention is not limited to the above-described specifically disclosed embodiments, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims.

[0088] 10 Estimation device 101 Input device 102 Display device 103 External I / F 103a Recording medium 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage device 108 Processor 109 Bus 201 Learning dataset acquisition unit 202 Segment division unit 203 Estimation unit 204 Model learning unit 205 Test data acquisition unit 206 Segment division unit 207 Estimation unit 208 Output unit 301 Learning dataset storage unit 302 Model parameter storage unit 303 Test data storage unit 304 Learned model parameter storage unit

Claims

1. A learning device comprising: a division unit that divides time series data included in given learning data into segments for each predetermined time interval; an estimator to which learning target parameters are set and an estimation unit that estimates, for each segment, parameters of a probability distribution to which correct output information included in the learning data follows, based on the segments; and a learning unit that learns the learning target parameters based on the parameters of the probability distribution for each segment and the correct output information, using a likelihood function related to a probability distribution expressed as the product of the probability distributions for each segment as a loss function, so as to minimize the loss function.

2. A learning device comprising: a division unit that divides time series data included in given learning data into segments for each predetermined time interval; an estimator to which a first learning target parameter is set, and an estimation unit that estimates, for each segment, parameters of a probability distribution to which true correct output information follows based on the segments; and a learning unit that learns second learning target parameters including the first learning target parameter based on the parameters of the probability distribution for each segment and observed output information included in the learning data, so as to maximize the posterior probability of the correct output information and the observed output information when the time series data and the first learning target parameter are given.

3. A learning method performed by a computer, comprising: a division procedure for dividing time series data included in given learning data into segments for each predetermined time interval; an estimation procedure for estimating, for each segment, parameters of a probability distribution to which correct output information included in the learning data follows, based on an estimator to which learning parameters are set and the segments; and a learning procedure for learning, based on the parameters of the probability distribution for each segment and the correct output information, a likelihood function related to a probability distribution expressed as the product of the probability distributions for each segment, so as to minimize the loss function.

4. A learning method performed by a computer, comprising: a division procedure for dividing time series data included in given learning data into segments for each predetermined time interval; an estimation procedure for estimating, for each segment, parameters of a probability distribution to which true correct output information follows, based on an estimator to which a first learning parameter is set and the segments; and a learning procedure for learning second learning parameters including the first learning parameter, based on the parameters of the probability distribution for each segment and observed output information included in the learning data, so as to maximize the posterior probability between the correct output information and the observed output information when the time series data and the first learning parameter are given.

Images