Information processing device, information processing method, and information processing program

Abstract

To enable effective classification of sequence data while suppressing an increase in time cost and calculation cost, an information processing device (1) comprises: an acquisition means (11) that acquires sequence data; a classification means (12) that performs classification processing for classifying the sequence data into any of a plurality of classes by using a score representing the certainty of classification; and a training means (13) that trains the classification means by using a parameter used for approximately performing likelihood ratio calculation and set according to the length of the sequence data.

Description

Information processing device, information processing method, and information processing program

[0001] The present invention relates to an information processing device, an information processing method, and an information processing program.

[0002] There is known a technique for classifying sequential data into one of a plurality of classes by referring to the sequential data. For example, Patent Literature 1 discloses an identification device that identifies to which of a plurality of classes, such as pre-registered people or other objects, a plurality of time-series data, which is an example of sequential data, belongs.

[0003] JP 2009-245314 A

[0004] Generally, in classification processes for classifying sequential data, it is necessary to sequentially calculate scores such as log-likelihoods used for classification. Therefore, in such classification processes or in the learning process of the classifier that performs the classification process, calculating scores by strict calculations imposes a heavy processing load. For this reason, in problems involving classification of sequential data, a method of approximately calculating the scores described above is often adopted.

[0005] As an example of an approximation method, a cutoff parameter is introduced, and a score is calculated based on the degree of approximation defined by the cutoff parameter. However, such a cutoff parameter is a hyperparameter that must be set in advance during learning. Therefore, in order to determine the optimal cutoff parameter value, conventionally, learning has been repeated while changing the value of the cutoff parameter, which has resulted in increased time and computational costs.

[0006] One aspect of the present invention has been made in view of the above-mentioned problems, and one objective thereof is to provide a technique that can effectively classify sequence data while suppressing increases in time and computational costs.

[0007] An information processing device according to one aspect of the present invention includes an acquisition means for acquiring sequence data, and a classification means for performing a classification process for classifying the sequence data into one of a plurality of classes, the classification process referring to a score representing the accuracy of the classification, the score being a likelihood ratio, the classification process including a process for classifying a certain data item in the sequence data into the class when the score corresponding to the certain data item reaches a threshold value of one of a plurality of classes, and updating the score using data included in the sequence data that follows the certain data item when the score corresponding to the certain data item does not reach any of the threshold values ​​of the plurality of classes, and a learning means for training the classification means using a parameter for approximately calculating the likelihood ratio, the parameter being set according to the length of the sequence data.

[0008] An information processing device according to one aspect of the present invention includes an acquisition means for acquiring sequence data; and a classification means for performing a classification process for classifying the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of the classification, wherein the score is a likelihood ratio. The classification process includes a process for classifying a certain data item in the sequence data into the class when the score corresponding to the certain data item reaches a threshold value of one of a plurality of classes, and updating the score using data included in the sequence data that follows the certain data item when the score corresponding to the certain data item does not reach any of the threshold values ​​of the plurality of classes. The classification means performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0009] An information processing method according to one aspect of the present invention includes, by at least one processor, acquiring sequence data and training a classifier that performs a classification process to classify the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of the classification, wherein the score is a likelihood ratio, and the classification process includes a process of classifying a certain data item in the sequence data into the class when the score corresponding to the certain data item reaches a threshold value of the certain data item, and updating the score using data included in the sequence data that follows the certain data item when the score corresponding to the certain data item does not reach any of the threshold values ​​of the certain data items, and the training step includes training the classifier using parameters for approximately calculating the likelihood ratio, the parameters being set according to the length of the sequence data.

[0010] An information processing method according to one aspect of the present invention includes, by at least one processor, acquiring sequence data, and using classification means to classify the sequence data into one of a plurality of classes, the classification process referring to a score representing the accuracy of the classification, wherein the score is a likelihood ratio, and the classification process includes classifying the sequence data into the class when a score corresponding to a certain data item in the sequence data reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain data item among the data included in the sequence data when the score corresponding to the certain data item does not reach any of the threshold values ​​of the plurality of classes, and the classification means performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0011] An information processing program according to one aspect of the present invention causes a computer to execute an acquisition process for acquiring sequential data; and a classification process for classifying the sequential data into one of a plurality of classes, the classification process referring to a score representing the accuracy of the classification, and training a classifier, wherein the score is a likelihood ratio, and the classification process includes a process for classifying a certain data item in the sequential data into the class when the score corresponding to the certain data item reaches a threshold value of one of a plurality of classes, and updating the score using a data item subsequent to the certain data item among the data items included in the sequential data when the score corresponding to the certain data item does not reach any of the threshold values ​​of the plurality of classes, and the training step causes the computer to execute a training process for training the classifier using a parameter for approximately calculating the likelihood ratio, the parameter being set according to the length of the sequential data.

[0012] An information processing program according to one aspect of the present invention causes a computer to execute an acquisition process for acquiring sequential data, and a classification process for classifying the sequential data into one of a plurality of classes using classification means, the classification process referring to a score representing the accuracy of the classification, wherein the score is a likelihood ratio, and the classification process includes a process for classifying a certain data item in the sequential data into the class when the score corresponding to the certain data item reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain data item among the data included in the sequential data when the score corresponding to the certain data item does not reach any of the threshold values ​​of the plurality of classes, and the classification means performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequential data.

[0013] According to one aspect of the present invention, it is possible to effectively classify sequence data while suppressing increases in time cost and calculation cost.

[0014] FIG. 1 is a block diagram showing the configuration of an information processing device according to exemplary embodiment 1. FIG. 2 is a flow diagram showing the flow of an information processing method according to exemplary embodiment 1. FIG. 3 is a block diagram showing the configuration of an information processing device according to exemplary embodiment 1. FIG. 4 is a flow diagram showing the flow of an information processing method according to exemplary embodiment 1. FIG. 5 is a block diagram showing the configuration of an information processing device according to exemplary embodiment 2. FIG. 6 is a flow diagram showing the flow of classification processing according to exemplary embodiment 2. FIG. 7 is a diagram for explaining classification processing according to exemplary embodiment 2. FIG. 8 is a flow diagram showing the flow of learning processing of the information processing device according to exemplary embodiment 2. FIG. 9 is a block diagram showing the configuration of an information processing device according to exemplary embodiment 3. FIG. 10 is a block diagram showing the configuration of a computer that functions as an information processing device according to each exemplary embodiment.

[0015] [First Exemplary Embodiment] A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described below.

[0016] Overview of Information Processing Device 1 The information processing device 1 according to this exemplary embodiment has a function of classifying sequence data input to an acquisition unit into one of a plurality of classes using a classification unit, and a function of training the classification unit. In other words, the information processing device 1 functions both as a classification device that classifies sequence data and as a learning device that trains the classification unit that performs the classification processing.

[0017] Here, the sequence data according to this exemplary embodiment is ordered data including one or more element data. In other words, the sequence data is composed of each element data. An example of the sequence data is video data or audio data. The sequence data may also be data representing trends in power consumption, etc. The sequence data is not limited to time-series data in which element data are arranged in chronological order, but may be any data in which the element data have some sort of sequential relationship. The element data is data that constitutes the sequence data, and an example is a frame in a video, but is not limited to this.

[0018] For example, the information processing device 1 calculates a score for classifying each element data included in the sequence data. A series of scores corresponding to the sequence data is constructed from a plurality of scores calculated for the sequence data.

[0019] Then, as an example, the information processing device 1 classifies the series data corresponding to the score series into one of multiple predetermined classes by comparing the scores included in the score series with one or more predetermined thresholds.

[0020] These processes executed by the information processing device 1 can be considered as a type of so-called SPRT (Sequential Probability Ratio Test), but are not limited thereto. Specific examples of classes do not limit the present exemplary embodiment, but include people, objects other than people, and people's actions.

[0021] As an example, the information processing device 1 is a device that acquires video images from a camera installed at the entrance of a building, an event venue, etc. as sequence data and classifies the subjects included in the video images into a plurality of pre-registered classes.

[0022] <Configuration of information processing device 1> Next, the configuration of the information processing device 1 will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the information processing device 1. As shown in Fig. 1, the information processing device 1 includes an acquisition unit 11, a classification unit 12, and a learning unit 13.

[0023] (Acquisition Unit 11) The acquisition unit 11 acquires sequence data. The acquisition unit 11 may acquire the sequence data all at once, or may acquire the sequence data sequentially. An example of the sequence data is as described above.

[0024] (Classification Unit 12) The classification unit 12 performs a classification process to classify the sequential data acquired by the acquisition unit 11 into one of a plurality of classes, using a score that indicates the accuracy of the classification. Here, a log-likelihood ratio can be given as an example of the score. Although the details of the classification process performed by the classification unit 12 do not limit the present exemplary embodiment, an example includes the following process, as described above: calculate a score from each element data included in the sequential data acquired by the acquisition unit 11; and refer to a score series obtained from the calculated scores, and compare the scores included in the score series with a predetermined threshold, thereby classifying the sequential data corresponding to the score series into one of a plurality of predetermined classes.

[0025] More specifically, the classification process executed by the classification unit 12 includes the following processes: if the score corresponding to a certain data item in the sequence data reaches the threshold value of one of a plurality of classes, classifying the sequence data into that class; if the score corresponding to the certain data item does not reach the threshold value of any of the plurality of classes, updating the score using the data item subsequent to the certain data item among the data items included in the sequence data.

[0026] (Learning Unit 13) The learning unit 13 causes the classifier 12 to learn using parameters (cutoff parameters, hyperparameters) for approximately calculating the likelihood ratio, which are set according to the length of the sequence data. While specific examples of the parameters do not limit the present exemplary embodiment, one example is a parameter N for approximately calculating the likelihood ratio as an N-th order Markov process.

[0027] In addition, when the length of the acquired sequence data is not constant, for example, the parameter may be set according to the maximum length of the sequence data. Furthermore, the learning unit 13 may set the parameter so that it has a positive correlation with the length of the sequence data acquired by the acquiring unit 11 (more specifically, the time series length), and may perform likelihood ratio calculation using the set parameter.

[0028] More specifically, as an example, the learning unit 13 may set the parameter N for approximating the Nth-order Markov process as follows: N-1=length of the sequence data.

[0029] Alternatively, the learning unit 13 may set the parameter N for approximating an Nth-order Markov process as N-1=length of the sequence data-M, where M is a value smaller than the sequence data. As an example, the magnitude of M may be set to about 10% of the length of the sequence data, or about 20% of the length of the sequence data.

[0030] The score series referred to by the learning unit 13 may be a score series of scores calculated by the classification unit 12 in the classification phase in which the information processing device 1 functions as a classification device, or may be a score series of scores calculated by the classification unit 12 by referring to series data acquired as learning data in the learning phase in which the information processing device 1 functions as a learning device.

[0031] As described above, the information processing device 1 according to this exemplary embodiment acquires sequence data and performs a classification process to classify the sequence data into one of a plurality of classes using a score that indicates the accuracy of the classification. The classification process includes a process of classifying a certain piece of data in the sequence data into the class if the score corresponding to the certain piece of data reaches a threshold value for one of a plurality of classes, and updating the score using the data subsequent to the certain piece of data included in the sequence data if the score corresponding to the certain piece of data does not reach any of the threshold values ​​for the plurality of classes. The learning unit 13 included in the information processing device 1 is configured to train the classifier 12 using parameters for approximately calculating a likelihood ratio, the parameters being set according to the length of the sequence data.

[0032] The inventors have found that by performing the above-described learning process, it is possible to acquire a log-likelihood ratio using any parameter value equal to or less than the set value of the parameter without performing the learning process again. For example, the inventors have found that by training the classifier 12 once by approximation using an N-th order Markov process using a parameter N set to N=T, it is possible to acquire a log-likelihood ratio using an N-th order Markov process (T-th order, T-1-th order, T-2-th order, ..., 2nd order, 1st order) using each value of N=T, T-1, T-2, ..., 2, 1.

[0033] Therefore, with the above configuration, it is possible to obtain log-likelihood ratios that are approximately calculated using various values ​​of parameters without repeating learning while changing the values ​​of parameters for approximation (cutoff parameters, hyperparameters). Therefore, it is possible to determine optimal parameter values ​​without repeating learning and perform classification processing using the determined parameters, thereby effectively classifying sequence data while suppressing increases in time cost and calculation cost.

[0034] <Flow of Information Processing Method S1> Next, the flow of the information processing method S1 according to the present exemplary embodiment 1 will be described with reference to Fig. 2. Fig. 2 is a flow diagram showing the flow of the information processing method S1. As shown in Fig. 2, in step S11, the acquisition unit 11 acquires sequence data.

[0035] In step S12, the learning unit 13 trains the classifying unit 12, which performs classification processing to classify the sequence data into one of a plurality of classes using a score that indicates the accuracy of the classification.

[0036] Here, the score is a likelihood ratio, and the classification process includes the steps of: classifying a certain piece of data in the sequence data into the class when the score corresponding to the certain piece of data in the sequence data reaches the threshold of one of a plurality of classes; and updating the score using the data subsequent to the certain piece of data included in the sequence data when the score corresponding to the certain piece of data does not reach any of the thresholds of the plurality of classes.

[0037] In addition, the step of training the classification unit 12 includes training the classification means using parameters for approximately calculating the likelihood ratio, which are set according to the length of the sequence data.

[0038] As described above, the information processing method S1 according to this exemplary embodiment acquires sequence data and classifies the sequence data into one of a plurality of classes, and includes a step of training the classifier 12 that performs the classification process using a score that represents the accuracy of the classification, where the score is a likelihood ratio. The classification process includes a step of classifying a certain piece of data in the sequence data into the class if the score corresponding to the certain piece of data reaches a threshold value of one of a plurality of classes, and updating the score using a piece of data included in the sequence data that follows the certain piece of data if the score corresponding to the certain piece of data does not reach any of the threshold values ​​of the plurality of classes. The step of training the classifier 12 includes training the classifier using parameters for approximately calculating the likelihood ratio, which are set according to the length of the sequence data.

[0039] According to the above configuration, it is possible to obtain log-likelihood ratios that are approximately calculated using various parameter values ​​without undergoing a process of repeating learning while changing the parameter values ​​for approximation, similar to the effect achieved by the information processing device 1. Therefore, it is possible to determine optimal parameter values ​​without repeating learning and perform classification processing using the determined parameters, thereby classifying sequence data while suppressing increases in time cost and calculation cost.

[0040] <Configuration of Information Processing Device 2> Next, the configuration of the information processing device 2 according to this exemplary embodiment will be described with reference to Fig. 3. Fig. 3 is a block diagram showing the configuration of the information processing device 2. As shown in Fig. 3, the information processing device 2 includes an acquisition unit 21 and a classification unit 22. On the other hand, the information processing device 2 does not include the learning unit included in the information processing device 1. However, the classification unit 22 included in the information processing device 2 performs classification processing using parameters for approximation determined through learning processing by the learning unit 13 included in the information processing device 1. The information processing device 2 according to this exemplary embodiment functions as a classification device that classifies sequence data.

[0041] (Acquisition Unit 21) The acquisition unit 21 acquires sequence data in the same manner as the acquisition unit 11 provided in the information processing device 1, which has already been described. The specific processing executed by the acquisition unit 21 is similar to that of the acquisition unit 11 provided in the information processing device 1, which has already been described, and therefore will not be described here.

[0042] (Classification Unit 22) The classification unit 22 performs classification processing to classify the sequence data into one of a plurality of classes, using a score that indicates the accuracy of the classification.

[0043] Here, the score is a likelihood ratio, and the classification process includes the steps of: classifying a certain piece of data in the sequence data into the class when the score corresponding to the certain piece of data in the sequence data reaches the threshold of one of a plurality of classes; and updating the score using the data subsequent to the certain piece of data included in the sequence data when the score corresponding to the certain piece of data does not reach any of the thresholds of the plurality of classes.

[0044] The classification unit 22 performs the classification process using parameters (cutoff parameters, hyperparameters) for approximately calculating the likelihood ratio, which are determined through a learning process using parameters set according to the length of the sequence data.

[0045] Here, the learning process is, for example, the same as the learning process executed by the learning unit 13 included in the information processing device 1. Furthermore, "parameters for approximately calculating the likelihood ratio, which are determined through a learning process using parameters set according to the length of the sequence data" refers to, for example, parameters determined through the following processes: a parameter N for approximately calculating the likelihood ratio, which is set according to the length of the sequence data, is used in a learning process to acquire log-likelihood ratios corresponding to each of a plurality of parameter values ​​equal to or less than the set value, and an evaluation process that refers to the acquired plurality of log-likelihood ratios to determine an optimal parameter value.

[0046] More specifically, as an example, the inventors have found that if the length of the sequence data is represented as T-1 and the classification unit 12 is trained once by an approximation using an Nth-order Markov process with a parameter N set to N=T, it is possible to obtain a log-likelihood ratio using an Nth-order Markov process (Tth, T-1th, T-2th, ..., 2nd, 1st) using each value of N=T, T-1, T-2, ..., 2, 1.

[0047] Then, by using a log-likelihood ratio using an approximation by a Markov process of each order (Tth order, T-1th order, T-2th order, ..., 2nd order, 1st order), it is possible to evaluate what kind of classification processing will be performed when a certain value of N is used. Then, it is possible to determine the value of N that will enable the most preferable classification processing. Here, the evaluation processing can be performed without repeating the learning processing while changing the value of N, so that the optimal value of the parameter N can be determined while suppressing information on time cost and calculation cost.

[0048] As described above, the information processing device 2 according to this exemplary embodiment acquires sequence data, and the classifier 12 performs a classification process of classifying the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of the classification, the score being a likelihood ratio. The classification process includes a process of classifying a certain piece of data in the sequence data into the class if the score corresponding to the certain piece of data reaches a threshold value for one of a plurality of classes, and updating the score using data included in the sequence data that follows the certain piece of data if the score corresponding to the certain piece of data does not reach any of the threshold values ​​for the plurality of classes. The classifier 12 performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0049] As described above, the classification unit 12 performs the classification process using parameters for approximating likelihood ratio calculations, which are determined through a learning process using parameters set according to the length of the sequence data.Therefore, the classification process can be suitably performed using parameters for approximation having preferred values ​​determined while suppressing information on time cost and calculation cost.

[0050] <Flow of Information Processing Method S2> Next, the flow of the information processing method S2 according to the present exemplary embodiment 1 will be described with reference to Fig. 4. Fig. 4 is a flow diagram showing the flow of the information processing method S2. As shown in Fig. 4, in step S21, the acquisition unit 21 acquires sequence data. In step S22, the classification unit 22 performs classification processing to classify the sequence data into one of a plurality of classes, using a score that indicates the accuracy of the classification.

[0051] Here, the score is a likelihood ratio, and the classification process includes the steps of: classifying a certain piece of data in the sequence data into the class when the score corresponding to the certain piece of data in the sequence data reaches the threshold of one of a plurality of classes; and updating the score using the data subsequent to the certain piece of data included in the sequence data when the score corresponding to the certain piece of data does not reach any of the thresholds of the plurality of classes.

[0052] The classification unit 22 performs the classification process using parameters for approximately calculating the likelihood ratio, which are determined through a learning process using parameters set according to the length of the sequence data.

[0053] Here, the learning process is, for example, the same as the learning process included in the information processing method S1. Also, the "parameters for approximately calculating the likelihood ratio, which are determined through a learning process using parameters set according to the length of the sequence data" refers to, for example, parameters determined through the following processes: (a) a parameter N for approximately calculating the likelihood ratio, which is set according to the length of the sequence data, is used in a learning process to acquire log-likelihood ratios corresponding to each of a plurality of parameter values ​​equal to or less than the set value; and (b) an evaluation process is performed with reference to the acquired log-likelihood ratios to determine an optimal parameter value.

[0054] According to the information processing method S2 configured as described above, similar to the effects achieved by the information processing device 2, classification processing can be preferably performed using parameters for approximation having preferred values ​​determined while suppressing information on time cost and calculation cost.

[0055]

[0033] A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first exemplary embodiment are designated by the same reference numerals, and their description will not be repeated.

[0056] 5 is a block diagram showing the configuration of the information processing device 1A according to this exemplary embodiment. The information processing device 1A includes a control unit 10A, a storage unit 20A, a communication unit 30A, and an input / output unit 40A.

[0057] (Communication Unit 30A) The communication unit 30A communicates with devices external to the information processing device 1A via a communication line. Although the specific configuration of the communication line does not limit this exemplary embodiment, examples of the communication line include a wireless local area network (LAN), a wired LAN, a wide area network (WAN), a public line network, a mobile data communication network, or a combination thereof. The communication unit 30A transmits data supplied from the control unit 10A to other devices, and supplies data received from other devices to the control unit 10A.

[0058] (Input / Output Unit 40A) Input / output devices such as a keyboard, a mouse, a display, a printer, and a touch panel are connected to the input / output unit 40A. The input / output unit 40A receives various types of information input to the information processing device 1A from the connected input devices. Furthermore, under the control of the control unit 10A, the input / output unit 40A outputs various types of information to connected output devices. Examples of the input / output unit 40A include an interface such as a USB (Universal Serial Bus).

[0059] (Acquisition Unit 11) The acquisition unit 11 acquires the sequence data SD. For example, the acquisition unit 11 acquires the sequence data SD from another device via the communication unit 30A. Alternatively, the acquisition unit 11 may acquire the sequence data SD input via the input / output unit 40A. Alternatively, the acquisition unit 11 may acquire the sequence data SD by reading it from the storage unit 20A or an externally connected storage device.

[0060] In this exemplary embodiment, the sequence data SD acquired by the acquisition unit 11 is, for example, time-series data having a sequence with respect to time, but this does not limit the exemplary embodiment.

[0061] In addition, in this exemplary embodiment, the series data SD acquired by the acquisition unit 11 is, as an example, video data including a plurality of frames, but this does not limit the present exemplary embodiment.

[0062] (Classification Unit 12) The classification unit 12 performs classification processing to classify the sequence data acquired by the acquisition unit 11 into one of a plurality of classes, using a score that indicates the accuracy of the classification. Here, the score is calculated by the classification unit 12, for example. A specific example of the score calculated by the classification unit 12 is not limited to this exemplary embodiment, but a log-likelihood ratio is one example.

[0063] More specifically, the classifier 12 approximately calculates the log-likelihood ratio by using a parameter (cutoff parameter, hyperparameter) N. Here, the specific approximation method does not limit the present exemplary embodiment, but an example is approximate calculation using an N-th order Markov process.

[0064] The classification unit 12 calculates the log likelihood ratio ^λ by an approximation calculation using an N-th order Markov process. ｋｌ (where ^λ represents λ with a hat) Here, represents sequence data whose length (time series length) is T, t is an index indicating time, k and l are indexes indicating classes, and ^p θ represents a conditional probability (also called likelihood or posterior density), and θ is a subscript representing a parameter that is updated by learning. As is clear from the above formula, the classification unit 12 classifies the sequence data into X （ｓ－Ｎ，ｓ） , X （ｓ－Ｎ，ｓー１） As shown in Fig. 1, by treating the log-likelihood ratio ^λ as an N-th order Markov process with parameter N for approximation, ｋｌ is calculated approximately.

[0065] Furthermore, the classification unit 12 calculates the log likelihood ratio ^λ by an approximation calculation using an N-th order Markov process. ｋｌ (where ^λ represents λ with a hat) In this example, as is clear from the above formula, the classification unit 12 also classifies the sequence data into X （ｔ－Ｎ，ｔ） As shown above, the log-likelihood ratio is calculated approximately by treating the process as an N-th order Markov process using a parameter N for approximation.

[0066] The log-likelihood ratio ^λ calculated as above ｋｌ The score matrix with the log-likelihood ratio as each element is defined by the following. Unless there is any particular confusion, the score matrix is ​​defined as ^λ ｋｌ The score matrix is, for example, defined as an antisymmetric matrix.

[0067] Note that the specific configuration used by the classification unit 12 to calculate the above probability (likelihood) does not limit this exemplary embodiment, but as an example, a deep learning algorithm using a neural network can be used.

[0068] The classification unit 12 also uses the score matrix ̂λ defined as above. ｋｌ From each component of The score sequence s defined by ｋ （１，ｔ） (sometimes simply referred to as score sequence s). Here, min{} is a function whose output value is the minimum value of the arguments. As is clear from the above explanation, the classification unit 12 calculates score sequences s for the number of classes. In other words, score sequences s for the number of classes are calculated from one piece of sequence data acquired by the acquisition unit 11.

[0069] (Flow of classification process by classification unit 12) When a score corresponding to a certain piece of data in the series data acquired by the acquisition unit 11 reaches the threshold value of one of multiple classes, the classification unit 12 classifies the series data into that class, and when the score corresponding to the certain piece of data does not reach the threshold values ​​of any of the multiple classes, the classification unit 12 updates the score using the data next to the certain piece of data included in the series data.

[0070] The specific flow of the classification process of the sequence data by the classification unit 12 will be described below with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of the classification process of the sequence data by the classification unit 12.

[0071] 6, in step S101, the classification unit 12 calculates a score from the sequence data. More specifically, the classification unit 12 calculates the score sequence defined in the above description.

[0072] (S102) Next, in step S102, the classification unit 12 determines whether any of the scores included in the score series has reached a predetermined threshold value Th. If any of the scores included in the score series has reached the predetermined threshold value Th, the process proceeds to step S103; if not, the process proceeds to step S104.

[0073] Here, the threshold Th may be set for each class. For example, a threshold Th0 to be referenced for classification into class 0, a threshold Th1 to be referenced for classification into class 1, and a threshold Th2 to be referenced for classification into class 2 may be set.

[0074] (S103) If any score in the score series reaches a predetermined threshold Th, in step S103, the classification unit 12 classifies the series data into a class corresponding to the score series including the score that reaches the threshold Th, and ends the classification process. The classification unit 12 stores the classification result in the storage unit 20A, for example.

[0075] (S104) On the other hand, if there is no score that has reached the predetermined threshold Th among the scores included in the score series, in step S104, the classification unit 12 updates the score and updates the score series using the updated score. After these update processes, the process of step S102 is performed.

[0076] 7 is a diagram for schematically explaining the results of the classification process by the classification unit 12. In the example of FIG. 7, the horizontal axis represents time, and the vertical axis represents the score value. In the example shown in FIG. 7, the score series s１ （１，ｔ） approaches the threshold value Th as time passes, and reaches the threshold value Th at time τ. This is because, in step S102 described above, the score series s at time τ １ （１，ｔ） reaches the threshold value, and in step S103, classification into class 1 is performed. On the other hand, as shown in FIG. 7, the score series s ０ （１，ｔ） , and the score sequence s indicating the classification into class 2 ２ （０，ｔ） does not approach the threshold value Th over time, and does not reach the threshold value Th at time τ. Therefore, in this example, the sequence data to be processed is not classified into either class 0 or class 2.

[0077] In this way, the classification unit 12 sequentially calculates score sequences for the number of classes from the acquired sequence data, and sequentially compares the calculated score sequences with thresholds to classify the sequence data into one of a plurality of classes. Note that the classification process described above can be considered as a type of SPRT (Sequential Probability Ratio Test) algorithm, but this term does not limit this exemplary embodiment.

[0078] (Learning unit 13) The learning unit 13 causes the classification unit 12 to learn using parameters for approximately calculating the likelihood ratio, which are set according to the length of the sequence data acquired by the acquisition unit 11.

[0079] Here, the specific example of the parameter does not limit the present exemplary embodiment, but an example is a parameter N for performing likelihood ratio calculations approximately as an Nth-order Markov process.

[0080] In addition, when the length of the acquired sequence data is not constant, for example, the parameter may be set according to the maximum length of the sequence data. Furthermore, the learning unit 13 may set the parameter so that it has a positive correlation with the length of the sequence data acquired by the acquiring unit 11 (more specifically, the time series length), and may perform likelihood ratio calculation using the set parameter.

[0081] More specifically, as an example, the learning unit 13 may set the parameter N for approximating the Nth-order Markov process as follows: N-1=length of the sequence data.

[0082] Alternatively, the learning unit 13 may set the parameter N for approximating an Nth-order Markov process as N-1=length of the sequence data-M, where M is a value smaller than the sequence data. As an example, the magnitude of M may be set to about 10% of the length of the sequence data, or about 20% of the length of the sequence data.

[0083] (Learning Process Flow by Learning Unit 13) FIG. 8 is a flowchart showing the learning process flow of the classification unit 12 by the learning unit 13. As shown in FIG.

[0084] (S201) In step S201, the learning unit 13 sets a parameter for approximately calculating the likelihood ratio, the parameter corresponding to the length of the sequence data acquired by the acquisition unit 11. More specifically, the learning unit 13 sets the parameter for approximately calculating the likelihood ratio to have a value corresponding to the length of the sequence data acquired by the acquisition unit 11. As an example, as described above, the learning unit 13 sets the parameter N for approximately calculating the likelihood ratio as an Nth-order Markov process as follows: N-1=length of the sequence data.

[0085] (S202) Subsequently, in step S202, the learning unit 13 trains the classifier 12 using the values ​​of the parameters for approximation set in step S201. As an example, the learning unit 13 calculates a log-likelihood ratio by approximation calculation using an N-th order Markov process using a parameter N having the value set in step S201, and calculates a loss function using the calculated log-likelihood ratio. The learning unit 13 then trains the classifier 12 by updating the training target parameters of the classifier 12 (if a neural network is used as the classifier 12, the adjustable parameters of the neural network) so as to increase or decrease the value of the loss function calculated as described above. In other words, the learning unit 13 trains the classifier 12 by updating the training target parameters of the classifier 12 so as to maximize or minimize the loss function calculated as described above. Note that, as a method for optimizing the training target parameters using the loss function, stochastic gradient descent, error propagation, etc. can be used, but are not limited to these.

[0086] Although the specific example of the loss function described above does not limit the present exemplary embodiment, as an example, a log-sum-exponential loss function L defined by the following formula: ＬＳＥＬ Alternatively, a log-sum-exponential loss function ^L defined by the following formula may be used. ＬＳＥＬ Here, Mk and Ik represent the sample size and index set of class k, respectively. That is, is.

[0087] Based on the inventor's findings based on the expression for the log-likelihood ratio, by having the learning unit 13 train the classifier 12 as described above, it becomes possible to acquire a log-likelihood ratio using any parameter value equal to or less than the set value of the parameter without performing the training process again. For example, by setting N=T using the sequence length T of the sequence data and training the classifier 12 once by approximation using an N-th order Markov process using the set parameter N, it is possible to acquire a log-likelihood ratio using an N-th order Markov process (T-th order, T-1-th order, T-2-th order, ..., 2nd order, 1st order) using each value of N=T, T-1, T-2, ..., 2, 1.

[0088] In other words, by having the learning unit 13 train the classification unit 12 as described above, a set of probabilities (posterior distribution) used to calculate the log-likelihood ratios can be obtained. Therefore, each of the log-likelihood ratios according to an N-th order Markov process (T-th order, T-1-th order, T-2-th order, ..., 2-th order, 1-th order) using these probabilities can be obtained without executing the learning process again.

[0089] In this way, the learning unit 13 performs the above-described learning, and thus a group of log-likelihood ratios can be calculated by one learning process.

[0090] (Evaluation unit 14) The evaluation unit 14 executes the classification process using, as the parameter value for the approximation, each of a plurality of values ​​equal to or less than the value used by the learning unit 13, and evaluates the execution result. As an example, when the learning unit 13 sets the parameter N for approximation to N=T using the sequence length T of the sequence data and causes the classification unit 12 to learn by N-th order Markov approximation using the set parameter N, the evaluation unit 14 can obtain, as the execution result of the classification process, a classification result by an N-th order Markov process (T-th order, T-1-th order, T-2-th order, ..., 2nd order, 1st order) using each value of N=T, T-1, T-2, ..., 2, 1.

[0091] The evaluation unit 14 evaluates each of the classification results by comparing them with a predetermined standard or by comparing them with each other. Here, specific examples of evaluation indices used by the evaluation unit 14 do not limit the present exemplary embodiment, but as an example, the evaluation may be performed using one or more indices selected from classification accuracy, time required for classification, calculation cost for classification, etc.

[0092] The evaluation unit 14 may also be configured to thin out a plurality of values ​​equal to or less than the value used by the learning unit 13 as the parameter value for the approximation, and then evaluate the execution result of the classification process using each of the thinned values. As an example, the evaluation unit 14 may thin out the values ​​of N by limiting N to odd numbers among the classification results of an N-th order Markov process (T-th order, T-1-th order, T-2-th order, ..., 2-th order, 1-th order) using each value of N = T, T-1, T-2, ..., 2, 1, and evaluate each of the classification results using each value of N N = , ..., 5, 3, 1 limited to the odd numbers. Furthermore, the value of N may be thinned out every 10, or the value of N may be thinned out by limiting the value of N to only values ​​smaller than a predetermined value.

[0093] The evaluation unit 14 also functions as a determination means for determining the values ​​of the parameters for the approximation by referring to the evaluation results obtained as described above. As an example, the evaluation unit 14 determines N, which has achieved the classification process with the highest classification accuracy, as the value for deployment (classification phase) N, from the classification results of an N-th order Markov process (T-th order, T-1-th order, T-2-th order, ..., 2-th order, 1-th order) using each value of N=T, T-1, T-2, ..., 2, 1. ｄｅｐ Then, the evaluation unit 14 determines the determined N ｄｅｐ is stored in the storage unit 20A.

[0094] As described above, the information processing device 1A employs a configuration in which the classifier 12 is trained using parameters for approximately calculating likelihood ratios, which are set according to the length of the sequence data acquired by the acquirer 11. With this configuration, as described above, it is possible to: acquire a set of probabilities and a set of log-likelihood ratios to be used in calculating log-likelihood ratios without repeating the learning process; and determine values ​​of the parameters for deployment (for the classification phase) by evaluating the classification results using these probabilities or log-likelihood ratios.

[0095] Therefore, the above configuration provides a technique that can effectively classify sequence data while suppressing increases in time costs and calculation costs.

[0096]

[0033] A third exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first and second exemplary embodiments are denoted by the same reference numerals, and their description will not be repeated.

[0097] 9 is a block diagram showing the configuration of the information processing device 2A according to this exemplary embodiment. As shown in FIG. 9, the information processing device 2A includes a control unit 10A, a storage unit 20A, a communication unit 30A, and an input / output unit 40A. The information processing device 2A functions as a classification device that classifies sequence data.

[0098] As shown in Fig. 9, the control unit 10A of the information processing device 2A includes an acquisition unit 11 and a classification unit 12, but does not include a learning unit 13 or an evaluation unit 14. The process performed by the acquisition unit 11 included in the control unit 10A of the information processing device 2A is similar to that performed by the acquisition unit 11 included in the information processing device 1A. Furthermore, the classification unit 12 of the information processing device 2A uses parameter values ​​for deployment (for the classification phase) determined by the evaluation unit 14 included in the information processing device 1A as parameter values ​​for approximation. As shown in Fig. 9, the parameter N ｄｅｐ is stored in the storage unit 20A as an example and is referenced by the classification unit 12.

[0099] Therefore, the classification unit 12 can be described as performing the classification process using a parameter (parameter Ndep) that is a parameter for approximately calculating the likelihood ratio and that is determined through a learning process using a parameter (parameter N) that is set according to the length of the sequence data acquired by the acquisition unit 11.

[0100] According to the information processing device 2A configured as described above, as described in detail in the second exemplary embodiment, the information processing device 2A executes the classification process using parameter values ​​for deployment (for the classification phase) determined through the following processes: (a) acquiring a set of probabilities and a set of log-likelihood ratios to be used in calculating the log-likelihood ratios without repeating the learning process; and (b) evaluating the classification results using these probabilities or log-likelihood ratios to determine the values ​​of the parameters for deployment (for the classification phase).

[0101] Therefore, the above configuration provides a technique that can effectively classify sequence data while suppressing increases in time costs and calculation costs.

[0102] [Example of implementation by software] The control blocks (particularly the classification units 12, 22, learning unit 13, and evaluation unit 14) of the information processing devices 1, 2, 1A, and 2A may be implemented by logic circuits (hardware) formed on an integrated circuit (IC chip) or the like, or may be implemented by software.

[0103] In the latter case, the information processing devices 1, 2, 1A, and 2A each include a computer that executes instructions from a program, which is software that realizes each function. This computer includes, for example, at least one processor (control device) and at least one computer-readable recording medium storing the program. The object of the present invention is achieved when the processor in the computer reads and executes the program from the recording medium. The processor may be, for example, a CPU (Central Processing Unit). The recording medium may be a "non-transitory tangible medium," such as a ROM (Read Only Memory), tape, disk, card, semiconductor memory, or programmable logic circuit. The computer may also include a RAM (Random Access Memory) for loading the program. The program may also be supplied to the computer via any transmission medium capable of transmitting the program (such as a communication network or broadcast waves). Note that one aspect of the present invention may also be realized in the form of a data signal embedded in a carrier wave, in which the program is embodied by electronic transmission.

[0104] [Additional Note 1] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiments are also included in the technical scope of the present invention.

[0105] [Additional Note 2] Part or all of the above-described embodiment can also be described as follows: However, the present invention is not limited to the following described aspects.

[0106] (Supplementary Note 1) An information processing device comprising: an acquiring means for acquiring sequence data; and a classifying means for performing a classification process for classifying the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of classification, wherein the score is a likelihood ratio, and the classification process includes a process for classifying the sequence data into a class when a score corresponding to a certain piece of data in the sequence data reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain piece of data included in the sequence data when the score corresponding to the certain piece of data does not reach any of the threshold values ​​of the plurality of classes, and the information processing device further comprises: a learning means for causing the classifying means to learn using a parameter for approximately calculating the likelihood ratio, the parameter being set according to the length of the sequence data.

[0107] (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the parameter is a parameter N for approximately performing the likelihood ratio calculation as an Nth-order Markov process.

[0108] (Supplementary Note 3) The information processing device according to Supplementary Note 2, wherein the classification means is trained using N set as N-1=length of the sequence data.

[0109] (Supplementary Note 4) The information processing device according to any one of Supplementary Notes 1 to 3, further comprising an evaluation means for executing the classification process using, as the parameter value, each of a plurality of values ​​equal to or less than the value used by the learning means, and evaluating the execution result.

[0110] (Supplementary Note 5) The information processing device according to Supplementary Note 4, wherein the evaluation means thins out a plurality of values ​​equal to or less than the value used by the learning means as the parameter value, and evaluates a result of execution of the classification process using each of the thinned values.

[0111] (Supplementary Note 6) The information processing device according to Supplementary Note 4 or 5, further comprising a determination unit that determines the value of the parameter by referring to the evaluation result by the evaluation unit.

[0112] (Supplementary Note 7) An information processing device comprising: an acquiring means for acquiring sequence data; and a classifying means for performing a classification process of classifying the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of classification, wherein the score is a likelihood ratio, and the classification process includes a process of classifying the sequence data into a class when a score corresponding to a certain data in the sequence data reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain data among data included in the sequence data when the score corresponding to the certain data does not reach any of the threshold values ​​of the plurality of classes, and the classifying means performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0113] (Supplementary Note 8) The information processing device according to Supplementary Note 7, wherein the parameter is a parameter N for approximately performing the likelihood ratio calculation as an Nth-order Markov process.

[0114] (Supplementary Note 9) The information processing device according to Supplementary Note 8, wherein the classification means is trained using N set as N-1=length of the sequence data.

[0115] (Supplementary Note 10) An information processing method, comprising: at least one processor acquiring sequence data; and training a classification means that performs a classification process of classifying the sequence data into one of a plurality of classes, the classification process referring to a score that represents the accuracy of the classification, wherein the score is a likelihood ratio, and the classification process includes a process of classifying the sequence data into a class when a score corresponding to a certain data in the sequence data reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain data among data included in the sequence data when the score corresponding to the certain data does not reach any of the threshold values ​​of the plurality of classes, and the training step includes training the classification means using parameters for approximately calculating a likelihood ratio, the parameters being set according to the length of the sequence data.

[0116] (Supplementary Note 11) An information processing method including: at least one processor acquiring sequence data; and using a classification means to classify the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of the classification, wherein the score is a likelihood ratio, and the classification process includes a process of classifying the sequence data into a class when a score corresponding to a certain data in the sequence data reaches a threshold of one of a plurality of classes, and updating the score using data subsequent to the certain data among data included in the sequence data when the score corresponding to the certain data does not reach any of the thresholds of the plurality of classes, and wherein the classification means performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0117] (Supplementary Note 12) An information processing program that causes a computer to execute an acquisition process of acquiring sequence data; and a classification process of classifying the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of the classification, and causes a classification means to train, wherein the score is a likelihood ratio, and the classification process includes a process of classifying the sequence data into a class when a score corresponding to a certain data in the sequence data reaches a threshold of one of a plurality of classes, and updating the score using data subsequent to the certain data among data included in the sequence data when the score corresponding to the certain data does not reach any of the thresholds of the plurality of classes, and in the training step, causes a computer to execute a training process of training the classification means using parameters for approximately calculating a likelihood ratio, the parameters being set according to the length of the sequence data.

[0118] (Supplementary Note 13) An information processing program that causes a computer to execute: an acquisition process that acquires sequence data; and a classification process that uses classification means to classify the sequence data into one of a plurality of classes, the classification process referring to a score that indicates the accuracy of the classification, wherein the score is a likelihood ratio, and the classification process includes a process of classifying a certain piece of data in the sequence data into the class when the score corresponding to the certain piece of data reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain piece of data included in the sequence data when the score corresponding to the certain piece of data does not reach any of the threshold values ​​of the plurality of classes, and the classification means performs the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0119] [Additional Note 3] Part or all of the above-described embodiment can also be expressed as follows.

[0120] (Supplementary Note 14) An information processing apparatus comprising at least one processor, the processor executing an acquisition process of acquiring sequence data; and a classification process of classifying the sequence data into one of a plurality of classes, the classification process referring to a score representing the accuracy of classification, and training a classification means, wherein the score is a likelihood ratio, the classification process includes a process of classifying a certain data in the sequence data into the class when the score corresponding to the certain data reaches a threshold value of one of a plurality of classes, and updating the score using data subsequent to the certain data among data included in the sequence data when the score corresponding to the certain data does not reach any of the threshold values ​​of the plurality of classes, and the training step executes a training process of training the classification means using parameters for approximately calculating a likelihood ratio, the parameters being set according to the length of the sequence data.

[0121] (Supplementary Note 15) An information processing apparatus comprising at least one processor, the processor executing: an acquisition process for acquiring sequence data; and a classification process for classifying the sequence data into one of a plurality of classes using classification means, the classification process referring to a score that indicates the accuracy of the classification, the score being a likelihood ratio, the classification process including: when a score corresponding to a certain piece of data in the sequence data reaches a threshold value of one of a plurality of classes, classifying the sequence data into the class, and when the score corresponding to the certain piece of data does not reach any of the threshold values ​​of the plurality of classes, updating the score using data subsequent to the certain piece of data included in the sequence data, the classification means performing the classification process using parameters for approximately calculating the likelihood ratio, the parameters being determined through a learning process using parameters set according to the length of the sequence data.

[0122] These information processing devices may further include a memory that stores a program for causing the processor to execute the acquisition process, the classification process, and the learning process. The program may also be recorded on a computer-readable, non-transitory, tangible recording medium.

[0123] REFERENCE SIGNS 1, 2, 1A, 2A... Information processing device 11, 21... Acquisition unit 12, 22... Classification unit 13... Learning unit 14... Evaluation unit 20A... Storage unit 30A... Communication unit 40A... Input / output unit

Claims

1. an acquisition means for acquiring sequence data; a classification means for performing a classification process of classifying the sequence data into one of a plurality of classes by referring to a score that indicates the accuracy of the classification; An information processing device comprising: The score is a likelihood ratio, The classification process includes: classifying the sequence data into a class when a score corresponding to a certain data item in the sequence data reaches a threshold value of one of a plurality of classes; and if the score corresponding to the certain data does not reach any of the thresholds of the plurality of classes, updating the score using data subsequent to the certain data among data included in the sequence data; The information processing device includes: The parameter is for approximately calculating the likelihood ratio, and the parameter is set according to the length of the sequence data, and a learning means is provided for causing the classification means to learn the parameter. Information processing device.

2. The parameter is a parameter N for approximately calculating the likelihood ratio as an N-th order Markov process. The information processing device according to claim 1 .

3. The classification means N-1 = length of the sequence data The training was performed using N set as The information processing device according to claim 2 .

4. The method further includes an evaluation unit that executes the classification process using each of a plurality of values ​​equal to or less than the value used by the learning unit as the parameter value, and evaluates the execution result. The information processing device according to claim 1 .

5. The evaluation means A plurality of values ​​equal to or less than the value used by the learning means are thinned out as the parameter values, and the execution result of the classification process using each of the thinned out values ​​is evaluated. The information processing device according to claim 4 .

6. The apparatus further includes a determination unit that determines the value of the parameter by referring to the evaluation result by the evaluation unit. The information processing device according to claim 4 .

7. an acquisition means for acquiring sequence data; a classification means for performing a classification process of classifying the sequence data into one of a plurality of classes by referring to a score that indicates the accuracy of the classification; Equipped with The score is a likelihood ratio, The classification process includes: classifying the sequence data into a class when a score corresponding to a certain data item in the sequence data reaches a threshold value of one of a plurality of classes; and if the score corresponding to the certain data does not reach any of the thresholds of the plurality of classes, updating the score using data subsequent to the certain data among data included in the sequence data; The classification means The classification process is performed using parameters for approximately calculating the likelihood ratio, which are determined through a learning process using parameters set according to the length of the sequence data. Information processing device.

8. The parameter is a parameter N for approximately calculating the likelihood ratio as an N-th order Markov process. The information processing device according to claim 7 .

9. At least one processor Obtaining sequence data; training a classification means that performs classification processing to classify the sequence data into one of a plurality of classes while referring to a score that indicates the accuracy of the classification; Including, The score is a likelihood ratio, The classification process includes: classifying the sequence data into a class when a score corresponding to a certain data item in the sequence data reaches a threshold value of one of a plurality of classes; and if the score corresponding to the certain data does not reach any of the thresholds of the plurality of classes, updating the score using data subsequent to the certain data among data included in the sequence data; The learning step includes: The method includes causing the classification means to learn using parameters for approximately calculating the likelihood ratio, the parameters being set according to the length of the sequence data. Information processing methods.

10. On the computer, an acquisition process for acquiring sequence data; training a classification means that performs classification processing to classify the sequence data into one of a plurality of classes while referring to a score that indicates the accuracy of the classification; Execute The score is a likelihood ratio, The classification process includes: classifying the sequence data into a class when a score corresponding to a certain data item in the sequence data reaches a threshold value of one of a plurality of classes; and if the score corresponding to the certain data does not reach any of the thresholds of the plurality of classes, updating the score using data subsequent to the certain data among data included in the sequence data; In the learning step, the computer A learning process is executed to train the classification means using parameters for approximately calculating likelihood ratios, the parameters being set according to the length of the sequence data. Information processing program.