WO2022137439A1 - 情報処理システム、情報処理方法、及びコンピュータプログラム - Google Patents

情報処理システム、情報処理方法、及びコンピュータプログラム Download PDF

Info

Publication number
WO2022137439A1
WO2022137439A1 PCT/JP2020/048472 JP2020048472W WO2022137439A1 WO 2022137439 A1 WO2022137439 A1 WO 2022137439A1 JP 2020048472 W JP2020048472 W JP 2020048472W WO 2022137439 A1 WO2022137439 A1 WO 2022137439A1
Authority
WO
WIPO (PCT)
Prior art keywords
likelihood ratio
likelihood
class
information processing
series data
Prior art date
Application number
PCT/JP2020/048472
Other languages
English (en)
French (fr)
Inventor
章記 海老原
大輝 宮川
Original Assignee
日本電気株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 日本電気株式会社 filed Critical 日本電気株式会社
Priority to PCT/JP2020/048472 priority Critical patent/WO2022137439A1/ja
Priority to US18/267,730 priority patent/US20240054400A1/en
Priority to JP2022570890A priority patent/JPWO2022137439A1/ja
Publication of WO2022137439A1 publication Critical patent/WO2022137439A1/ja

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • G06N20/10Machine learning using kernel methods, e.g. support vector machines [SVM]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent

Definitions

  • This disclosure relates to, for example, the technical fields of information processing systems, information processing methods, and computer programs that process information related to classification.
  • Patent Document 1 discloses learning a support vector machine used for determining a likelihood.
  • Patent Document 2 discloses that a support vector machine or logistic regression may be used for learning a discriminator using the likelihood.
  • Patent Document 3 in a device for determining whether or not a person included in an image is a registered person, there is a technique for performing folder classification processing on an image file based on predetermined recognition conditions and predominant conditions. It has been disclosed.
  • This disclosure is intended to improve the related techniques mentioned above.
  • One aspect of the information processing system of this disclosure is an acquisition means for acquiring a plurality of elements included in the series data, and a class to which the series data belongs based on at least two consecutive elements among the plurality of elements.
  • the calculation means for calculating the likelihood ratio indicating the likelihood the classification means for classifying the series data into at least one class among a plurality of classes that are classification candidates based on the likelihood ratio, and the series data Using a loss function, the likelihood ratio is large when the correct class to which it belongs is in the molecule of the likelihood ratio, and the likelihood ratio is small when the correct class is in the denominator of the likelihood ratio. It is provided with a learning means for learning about the calculation of the likelihood ratio.
  • One aspect of the information processing method of this disclosure is to acquire a plurality of elements included in the series data and to determine the likelihood of the class to which the series data belongs based on at least two consecutive elements among the plurality of elements.
  • the indicated likelihood ratio is calculated, and the series data is classified into at least one class among a plurality of classification candidate classes based on the likelihood ratio, and the correct answer class to which the series data belongs is the likelihood ratio.
  • Learning about the calculation of the likelihood ratio is performed using a loss function in which the likelihood ratio becomes large when it is in a molecule and the likelihood ratio becomes small when the correct answer class is in the denominator of the likelihood ratio. ..
  • One aspect of the computer program of this disclosure is to acquire a plurality of elements contained in the series data and show the likelihood of the class to which the series data belongs based on at least two consecutive elements among the plurality of elements.
  • the likelihood ratio is calculated, and based on the likelihood ratio, the series data is classified into at least one class among a plurality of classes that are candidates for classification, and the correct answer class to which the series data belongs is the molecule of the likelihood ratio.
  • the likelihood ratio is large, and when the correct answer class is in the denominator of the likelihood ratio, the likelihood ratio is small. Operate the computer.
  • FIG. 1 is a block diagram showing a hardware configuration of an information processing system according to the first embodiment.
  • the information processing system 1 includes a processor 11, a RAM (Random Access Memory) 12, a ROM (Read Only Memory) 13, and a storage device 14.
  • the information processing system 1 may further include an input device 15 and an output device 16.
  • the processor 11, the RAM 12, the ROM 13, the storage device 14, the input device 15, and the output device 16 are connected via the data bus 17.
  • Processor 11 reads a computer program.
  • the processor 11 is configured to read a computer program stored in at least one of the RAM 12, the ROM 13, and the storage device 14.
  • the processor 11 may read a computer program stored in a computer-readable recording medium using a recording medium reading device (not shown).
  • the processor 11 may acquire (that is, read) a computer program from a device (not shown) located outside the information processing system 1 via a network interface.
  • the processor 11 controls the RAM 12, the storage device 14, the input device 15, and the output device 16 by executing the read computer program.
  • a functional block for executing a classification using a likelihood ratio and a learning process related thereto is realized in the processor 11.
  • processor 11 a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an FPGA (field-programmable get array), a DSP (Demand-Side Platform), an ASIC Circuit, etc.
  • CPU Central Processing Unit
  • GPU Graphics Processing Unit
  • FPGA field-programmable get array
  • DSP Demand-Side Platform
  • ASIC Circuit etc.
  • processor 11 one of the above-mentioned examples may be used, or a plurality of processors 11 may be used in parallel.
  • the RAM 12 temporarily stores the computer program executed by the processor 11.
  • the RAM 12 temporarily stores data temporarily used by the processor 11 while the processor 11 is executing a computer program.
  • the RAM 12 may be, for example, a D-RAM (Dynamic RAM).
  • the ROM 13 stores a computer program executed by the processor 11.
  • the ROM 13 may also store fixed data.
  • the ROM 13 may be, for example, a P-ROM (Programmable ROM).
  • the storage device 14 stores data stored in the information processing system 1 for a long period of time.
  • the storage device 14 may operate as a temporary storage device of the processor 11.
  • the storage device 14 may include, for example, at least one of a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device.
  • the input device 15 is a device that receives an input instruction from the user of the information processing system 1.
  • the input device 15 may include, for example, at least one of a keyboard, a mouse and a touch panel.
  • the input device 15 may be a dedicated controller (operation terminal). Further, the input device 15 may include a terminal owned by the user (for example, a smartphone, a tablet terminal, or the like).
  • the input device 15 may be a device capable of voice input including, for example, a microphone.
  • the output device 16 is a device that outputs information about the information processing system 1 to the outside.
  • the output device 16 may be a display device (for example, a display) capable of displaying information about the information processing system 1.
  • the display device here may be a television monitor, a personal computer monitor, a smartphone monitor, a tablet terminal monitor, or another mobile terminal monitor.
  • the display device may be a large monitor, a digital signage, or the like installed in various facilities such as a store.
  • the output device 16 may be a device that outputs information in a format other than an image.
  • the output device 16 may be a speaker that outputs information about the information processing system 1 by voice.
  • FIG. 2 is a block diagram showing a functional configuration of the information processing system according to the first embodiment.
  • the information processing system 1 includes a classification device 10 and a learning unit 300.
  • the classification device 10 is a device that classifies the input series data into classes, and has a data acquisition unit 50, a likelihood ratio calculation unit 100, and a class classification unit 200 as processing blocks for realizing the function. It is configured in preparation.
  • the learning unit 300 is configured to be able to execute the learning process related to the classification device 10.
  • the classification device 10 may be configured to include the learning unit 300.
  • Each of the data acquisition unit 50, the likelihood ratio calculation unit 100, the classification unit 200, and the learning unit 300 may be realized by the processor 11 (see FIG. 1) described above.
  • the data acquisition unit 50 is configured to be able to acquire a plurality of elements included in the series data.
  • the data acquisition unit 50 may acquire data directly from an arbitrary data acquisition device (for example, a camera, a microphone, etc.), or may read data previously acquired by the data acquisition device and stored in the storage or the like. It may be a thing.
  • the data acquisition unit 50 may be configured to acquire data from each of the plurality of cameras.
  • the elements of the series data acquired by the data acquisition unit 50 are configured to be output to the likelihood ratio calculation unit 100.
  • the series data is data including a plurality of elements arranged in a predetermined order, and for example, time series data can be mentioned as an example. More specific examples of series data include, but are not limited to, video data and audio data.
  • the likelihood ratio calculation unit 100 is configured to be able to calculate the likelihood ratio based on at least two consecutive elements among the plurality of elements acquired by the data acquisition unit 50.
  • the "likelihood ratio" here is an index showing the likelihood of the class to which the series data belongs. Specific examples and specific calculation methods of the likelihood ratio will be described in detail in other embodiments described later.
  • the class classification unit 200 is configured to be able to classify series data based on the likelihood ratio calculated by the likelihood ratio calculation unit 100.
  • the class classification unit 200 selects at least one class to which the series data belongs from a plurality of classes that are classification candidates.
  • the plurality of classes that are candidates for classification may be preset. Alternatively, the plurality of classes that are classification candidates may be appropriately set by the user, or may be appropriately set based on the type of series data to be handled and the like.
  • the learning unit 300 learns about the calculation of the likelihood ratio using the loss function. Specifically, learning related to the calculation of the likelihood ratio is executed so that the classification based on the likelihood ratio is performed accurately.
  • the loss function used by the learning unit 300 has a large likelihood ratio when the correct answer class to which the series data belongs is in the numerator of the likelihood ratio, and a small likelihood ratio when the correct answer class is in the denominator of the likelihood ratio. It is defined as a function.
  • the loss function may be preset as a function satisfying such a definition. Specific examples of the loss function will be described in detail in other embodiments described later.
  • FIG. 3 is a flowchart showing the operation flow of the classification device in the information processing system according to the first embodiment.
  • the data acquisition unit 50 first acquires the elements included in the series data (step S11).
  • the data acquisition unit 50 outputs the elements of the acquired series data to the likelihood ratio calculation unit 100.
  • the likelihood ratio calculation unit 100 calculates the likelihood ratio based on the acquired two or more elements (step S12).
  • the class classification unit 200 classifies the class based on the calculated likelihood ratio (step S13).
  • the classification may determine one class to which the series data belongs, or may determine a plurality of classes to which the series data is likely to belong.
  • the class classification unit 200 may output the result of the class classification to a display or the like. Further, the class classification unit 200 may output the result of the class classification by voice via a speaker or the like.
  • FIG. 4 is a flowchart showing the flow of operation of the learning unit in the information processing system according to the first embodiment.
  • the training data is first input to the learning unit 300 (step S101).
  • the training data may be configured, for example, as a set of sequence data and information about the correct answer class to which the sequence data belongs (that is, correct answer data).
  • the learning unit 300 calculates the loss function using the input training data (step S102).
  • the loss function here has a large likelihood ratio when the correct answer class to which the series data belongs is in the numerator of the likelihood ratio, and the likelihood is when the correct answer class is in the denominator of the likelihood ratio. It is a function that reduces the ratio.
  • the learning unit 300 adjusts the parameters (specifically, the parameters of the model for calculating the likelihood ratio) so that the calculated loss function becomes smaller (step S103). That is, the learning unit 300 optimizes the parameters of the model for calculating the likelihood ratio.
  • the learning unit 300 optimizes the parameters of the model for calculating the likelihood ratio.
  • a parameter optimization method using a loss function it is possible to appropriately adopt an existing technique.
  • An example of the optimization method is the backpropagation method, but other methods may be used.
  • the learning unit 300 determines whether or not all learning has been completed (step S104).
  • the learning unit 300 may determine whether or not all the learning has been completed, for example, depending on whether or not all the training data have been input.
  • the learning unit 300 may determine whether or not all learning has been completed, depending on whether or not a predetermined period has elapsed from the start of learning.
  • the learning unit 300 may determine whether or not all the learning has been completed depending on whether or not the processes of steps S101 to S103 described above have been looped a predetermined number of times.
  • step S104 When it is determined that all learning is completed (step S104: YES), a series of processes is completed. On the other hand, when it is determined that all the learning has not been completed (step S104: NO), the learning unit 300 starts the process again from step S101. As a result, the learning process using the training data is repeated, and the parameters are adjusted to more optimum ones.
  • the learning unit 300 executes learning related to the calculation of the likelihood ratio used for the classification.
  • the likelihood ratio becomes large especially when the correct answer class to which the series data belongs is in the numerator of the likelihood ratio, and the likelihood ratio becomes small when the correct answer class is in the denominator of the likelihood ratio. Learning is done using. By using such a loss function, learning can be performed so that the penalty when the class is incorrect is large and the penalty when the class is correct is small. As a result, it becomes possible to appropriately select at least one class to which the series data belongs from a plurality of classes that are classification candidates.
  • the information processing system 1 according to the second embodiment will be described with reference to FIGS. 5 and 6.
  • the second embodiment differs from the first embodiment described above in only a part of the operation.
  • the device configuration (see FIGS. 1 and 2) and the operation of the classification device 10 (see FIG. 3). ) And the like may be the same as in the first embodiment. Therefore, in the following, the parts different from the first embodiment will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 5 is a flowchart showing an operation flow of the information processing system according to the second embodiment.
  • the same reference numerals are given to the same processes as those shown in FIG.
  • the training data is first input to the learning unit 300 (step S101).
  • the learning unit 300 calculates the loss function using the input training data, but in the second embodiment, in particular, among the N classes (where N is a natural number) which are the classification candidates of the series data. , Calculate the loss function considering the likelihood ratio of the N ⁇ (N-1) pattern, whose denominator is the likelihood indicating the likelihood belonging to one class and the likelihood indicating the likelihood belonging to another class is the molecule. (Step S201). Similar to the first embodiment, this loss function also has a large likelihood ratio when the correct answer class to which the series data belongs is in the numerator of the likelihood ratio, and the likelihood is also when the correct answer class is in the denominator of the likelihood ratio. It is a function that reduces the ratio. The likelihood ratio considered in the loss function will be described in detail later with a specific example.
  • the learning unit 300 adjusts the parameters so that the calculated loss function becomes smaller (step S103). That is, the learning unit 300 optimizes the parameters of the model for calculating the likelihood ratio.
  • the learning unit 300 determines whether or not all learning has been completed (step S104). When it is determined that all learning is completed (step S104: YES), a series of processes is completed. On the other hand, when it is determined that all the learning has not been completed (step S104: NO), the learning unit 300 starts the process again from step S101.
  • FIG. 6 is a matrix diagram showing an example of the likelihood ratio considered by the learning unit in the information processing system according to the second embodiment.
  • the likelihood ratio on the diagonal of the matrix (the likelihood ratio shaded in gray in FIG. 6) has the same likelihood for the denominator and the numerator. Specifically, log ⁇ p (X
  • y 0) / p (X
  • y 0) ⁇ in the first row from the top and the first column from the left, the second row from the top, and the second column from the left. log ⁇ p (X
  • y 1) / p (X
  • y 1) ⁇ , 3rd row from the top, 3rd column from the left log ⁇ p (X
  • y 2) / p (X
  • y 2) ⁇ has the same denominator and molecule, respectively.
  • the likelihood ratios at positions facing each other with the likelihood ratio on the diagonal line opposite to each other have opposite denominators and numerator. Specifically, log ⁇ p (X
  • y 0) / p (X
  • y 1) ⁇ in the first row from the top and the second column from the left, the second row from the top, and the first column from the left. In the log ⁇ p (X
  • y 1) / p (X
  • y 0) ⁇ , the denominator and the molecule are reversed.
  • the likelihood ratios at which the denominator and numerator on the diagonal are the same are all log1, and their values are zero. Therefore, the likelihood ratio at which the numerator is the same as the denominator on the diagonal line is a value that is practically meaningless even when considered in terms of the loss function. Therefore, the likelihood ratio at which the denominator and numerator on the diagonal are the same is not considered in the loss function.
  • the number of remaining likelihood ratios excluding the diagonal likelihood ratios is N ⁇ (N-1), where N is the number of classes.
  • the likelihood ratios of these N ⁇ (N-1) patterns (that is, the likelihood ratios excluding the diagonal likelihood ratios in the matrix) are considered in the loss function.
  • a specific example of the loss function considering the likelihood ratio of the N ⁇ (N-1) pattern will be described in detail in another embodiment described later.
  • the likelihood indicating the likelihood of belonging to one class is used as the denominator, and the likelihood indicating the likelihood of belonging to another class is defined as the denominator.
  • Learning is performed using a loss function that takes into account the likelihood ratio of the N ⁇ (N-1) pattern as the numerator.
  • the information processing system 1 according to the third embodiment will be described with reference to FIG. 7. It should be noted that the third embodiment may be the same as the first and second embodiments except for a part of the operation different from the above-mentioned first and second embodiments. Therefore, in the following, the parts different from each of the above-described embodiments will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 7 is a flowchart showing an operation flow of the information processing system according to the third embodiment.
  • the same reference numerals are given to the same processes as those shown in FIG.
  • the training data is first input to the learning unit 300 (step S101).
  • the learning unit 300 calculates the loss function using the input training data, but in the third embodiment, it is possible that the learning unit 300 belongs to one of the N classes that are candidates for classification of the series data. Calculate a loss function considering a part of the likelihood ratio of the N ⁇ (N-1) pattern, whose denominator is the likelihood indicating the likelihood and the likelihood indicating the likelihood belonging to another class is the molecule (step S301). ). That is, the learning unit 300 according to the third embodiment does not consider all the likelihood ratios of the N ⁇ (N-1) patterns described in the second embodiment, but considers only a part thereof.
  • this loss function also has a large likelihood ratio when the correct answer class to which the series data belongs is in the numerator of the likelihood ratio, and when the correct answer class is in the denominator of the likelihood ratio. It is a function that reduces the likelihood ratio.
  • a specific example of the loss function considering the likelihood ratio of a part of the N ⁇ (N-1) pattern will be described in detail in other embodiments described later.
  • step S104 determines whether or not all learning has been completed.
  • step S104 determines whether or not all learning has been completed.
  • a part of the likelihood ratios to be considered in the loss function may be selected in advance by the user or the like, or may be automatically selected by the learning unit 300. You may do it.
  • the learning unit 300 selects a part of the likelihood ratios to be considered in the loss function, the learning unit 300 may select the likelihood ratios according to a predetermined rule set in advance. Alternatively, the learning unit 300 may determine whether or not to select based on the calculated likelihood ratio value.
  • a method of selecting only the likelihood ratio of one row or one column in the matrix shown in FIG. 6 can be mentioned.
  • the likelihood ratio to be considered in the loss function only the likelihood ratio in the first row of the matrix shown in FIG. 6 may be selected, or only the likelihood ratio in the second row may be selected. Only the likelihood ratio on the third line may be selected.
  • only the likelihood ratio of the first column of the matrix may be selected, only the likelihood ratio of the second column may be selected, or only the likelihood ratio of the third column may be selected. ..
  • only the likelihood ratio of some multiple rows or some multiple columns in the matrix may be selected. Specifically, only the likelihood ratios of the first and second rows of the matrix may be selected, only the likelihood ratios of the second and third rows may be selected, and the third row may be selected. And only the likelihood ratio in the first line may be selected. Alternatively, only the likelihood ratios of the first and second columns of the matrix may be selected, only the likelihood ratios of the second and third columns may be selected, and only the third and first columns may be selected. Only the eye likelihood ratio may be selected.
  • likelihood ratio selection example is merely an example, and other likelihood ratios may be selected as the likelihood ratios considered in the loss function.
  • the likelihood ratio considered for the loss function may be randomly selected regardless of the row or column.
  • the likelihood indicating the likelihood of belonging to one class is used as the denominator, and the likelihood indicating the likelihood of belonging to another class is used as the numerator.
  • Learning is performed using a loss function that considers a part of the likelihood ratios of the N ⁇ (N-1) patterns.
  • a loss function that considers a part of the likelihood ratios of the N ⁇ (N-1) patterns.
  • the third embodiment in particular, by appropriately selecting the likelihood ratio to be considered for the loss function from the N ⁇ (N-1) patterns, all the likelihood ratios of the N ⁇ (N-1) patterns are selected. Compared to the case of considering, it is possible to perform efficient learning. For example, the learning efficiency can be improved by selecting only the likelihood ratio having a relatively large influence on the loss function and not selecting the likelihood ratio having a relatively small influence on the loss function.
  • the information processing system 1 according to the fourth embodiment will be described with reference to FIGS. 8 and 9.
  • the fourth embodiment describes a specific selection example of the third embodiment described above (that is, a selection example of a part of the likelihood ratios considered in the loss function), and the other parts are described. , It may be the same as the third embodiment. Therefore, in the following, the parts different from each of the above-described embodiments will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 8 is a flowchart showing an operation flow of the information processing system according to the fourth embodiment.
  • the same reference numerals are given to the same processes as those shown in FIG.
  • the training data is first input to the learning unit 300 (step S101).
  • the learning unit 300 calculates the loss function using the input training data, but in the fourth embodiment, in particular, among the likelihood ratios of the N ⁇ (N-1) patterns already described, the correct answer class.
  • Calculates a loss function in consideration of the likelihood ratio in the numerator step S401). That is, the learning unit 300 according to the fourth embodiment selects the likelihood ratio in which the correct answer class is in the molecule as the likelihood ratio of a part of the N ⁇ (N-1) pattern described in the third embodiment.
  • this loss function also has a large likelihood ratio when the correct answer class to which the series data belongs is in the numerator of the likelihood ratio, and when the correct answer class is in the denominator of the likelihood ratio. It is a function that reduces the likelihood ratio. Specific examples of the loss function considering the likelihood ratio in which the correct answer class is in the numerator will be described in detail in other embodiments described later.
  • step S104 determines whether or not all learning has been completed.
  • step S104 determines whether or not all learning has been completed.
  • FIG. 9 is a matrix diagram showing an example of the likelihood ratio considered by the learning unit in the information processing system according to the fourth embodiment.
  • the likelihood ratios are arranged like a skew-symmetric matrix.
  • the learning unit 300 according to the fourth embodiment has a likelihood ratio in which the correct answer class is in the numerator from the likelihood ratios of the N ⁇ (N-1) pattern excluding the diagonal likelihood ratios in such a matrix. Select to consider the loss function.
  • the learning unit 300 selects the likelihood ratio in which class 1 is in the molecule from the likelihood ratios of the N ⁇ (N-1) pattern and considers it in the loss function. Specifically, only the likelihood ratio (excluding the diagonal likelihood ratio) in the second row from the top of FIG. 9 is selected and considered in the loss function. In this case, the log ⁇ p (X
  • y 1) / p (X
  • y 0) ⁇ in the second row from the top and the first column from the left, and the log in the second row from the top and the third column from the left. ⁇ P (X
  • y 1) / p (X
  • y 2) ⁇ will be considered in the loss function. That is, the gray and unshaded likelihood ratios in FIG. 9 are taken into account in the loss function.
  • the learning unit 300 is likely that class 0 is in the numerator from the likelihood ratios of the N ⁇ (N-1) patterns.
  • the degree ratio may be selected and taken into account in the loss function. Specifically, only the likelihood ratio (excluding the diagonal likelihood ratio) in the first row from the top of FIG. 9 may be selected and considered in the loss function. In this case, the log ⁇ p (X
  • y 0) / p (X
  • y 1) ⁇ in the first row from the top and the second column from the left, and the log in the first row from the top and the third column from the left. ⁇ P (X
  • y 0) / p (X
  • y 2) ⁇ will be considered in the loss function.
  • the learning unit 300 has class 2 in the numerator from the likelihood ratios of the N ⁇ (N-1) pattern.
  • the likelihood ratio may be selected and taken into account in the loss function. Specifically, only the likelihood ratio (excluding the diagonal likelihood ratio) in the third row from the top of FIG. 9 may be selected and considered in the loss function. In this case, the log ⁇ p (X
  • y 2) / p (X
  • y 0) ⁇ in the third row from the top and the first column from the left, and the log in the third row from the top and the second column from the left. ⁇ P (X
  • y 2) / p (X
  • y 1) ⁇ will be considered in the loss function.
  • a loss function considering the likelihood ratio in which the correct answer class is in the molecule among the N ⁇ (N-1) patterns is used. Learning is executed. If such a loss function is used, appropriate learning will be performed as in each of the above-described embodiments. Therefore, at least one class to which the series data belongs is appropriately selected from a plurality of classes that are classification candidates. It is possible to select to. Further, in the fourth embodiment, in particular, the likelihood ratio in which the correct answer class is in the numerator (in other words, the likelihood ratio that may have a great influence on the loss function) is considered in the loss function, so that N ⁇ (N). -1) Efficient learning can be performed as compared with the case where all the likelihood ratios of the pattern are considered.
  • the information processing system 1 according to the fifth embodiment will be described with reference to FIG.
  • the fifth embodiment describes specific examples of the loss function used in the first to fourth embodiments described above, and the apparatus configuration and operation flow are the same as those of the first to fourth embodiments. May be. Therefore, in the following, the parts different from each of the above-described embodiments will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 10 is a graph showing an example of a sigmoid function used in the information processing system according to the fifth embodiment.
  • the learning unit 300 performs learning using a loss function including a sigmoid function. More specifically, the learning unit 300 performs learning using a loss function including a sigmoid function as a non-linear function acting on the likelihood ratio.
  • a loss function including a sigmoid function as a non-linear function acting on the likelihood ratio.
  • K is the number of classes
  • M is the number of data
  • T is the time series length.
  • k is a subscript in the row direction and l is a subscript in the column direction (that is, a subscript indicating the row number and the column number in the matrix shown in FIG. 6 or the like).
  • is the Kronecker delta, which is "1" when the subscripts match, and "0" in other cases.
  • is a likelihood ratio, and in the above formula (1), it represents the log-likelihood ratio of the k-th row and l-th column at time t.
  • is a sigmoid function and is included as a nonlinear function acting on the likelihood ratio ⁇ .
  • 1 / MT in equation (1) is an average for all data and time series direction.
  • 1 / 2K is the product of 1 / K, which is the average of K rows, and 1/2, which is the average of the skew-symmetric matrix.
  • 1 / (K-1) is the average of the (K-1) column obtained by subtracting one diagonal column from the K column.
  • the Kronecker delta becomes "1"
  • the k line is the incorrect answer class (that is, a class other than the correct answer class)
  • the Kronecker ⁇ Is "0"
  • the value of the likelihood ratio becomes large.
  • row k is an incorrect answer class (in other words, if the incorrect answer class is in the numerator of the likelihood ratio)
  • the value of the likelihood ratio becomes small.
  • Equation (2) can be mentioned as a loss function that considers only the likelihood ratio in which the correct answer class is in the numerator.
  • K is the number of classes
  • M is the number of data
  • T is the time series length.
  • k is a subscript in the row direction
  • l is a subscript in the column direction (that is, a subscript indicating the row number and the column number in the matrix shown in FIG. 6 or the like).
  • is a likelihood ratio, and in the above formula (1), it represents the log-likelihood ratio of the k-th row and l-th column at time t.
  • is a sigmoid function and is included as a nonlinear function acting on the likelihood ratio ⁇ .
  • equation (2) Since equation (2) considers only the rows of the correct answer class, it includes a step of taking the sum of K rows and averaging the K rows and the skew-symmetric matrix at 1 / 2K as compared with the equation (1) already described. Not done. Further, the ⁇ portion of the Kronecker in the equation (1) is “1”.
  • the learning unit 300 uses a loss function including a sigmoid function ⁇ .
  • a loss function including a sigmoid function ⁇ .
  • the loss functions (that is, equations (1) and (2)) given in the fifth embodiment are examples, and different loss functions may be created by using the sigmoid function. Further, the loss function may be created by using another nonlinear function instead of the sigmoid function. For example, a loss function including a logistic function may be used as in the embodiment described later.
  • FIG. 11 is a graph showing an example of a logistic function used in the information processing system according to the sixth embodiment.
  • the learning unit 300 performs learning using a loss function including a logistic function. More specifically, the learning unit 300 performs learning using a loss function including a logistic function as a non-linear function acting on the likelihood ratio.
  • a loss function including a logistic function as a non-linear function acting on the likelihood ratio.
  • K is the number of classes
  • M is the number of data
  • T is the time series length.
  • k is a subscript in the row direction and l is a subscript in the column direction (that is, a subscript indicating the row number and the column number in the matrix shown in FIG. 6 or the like).
  • is the Kronecker delta, which is "1" when the subscripts match, and "0" in other cases.
  • is a likelihood ratio, and in the above formula (3), it represents the log-likelihood ratio of the k-th row and l-th column at time t.
  • Logistic is a logistic function and is included as a non-linear function acting on the likelihood ratio ⁇ .
  • 1 / MT in equation (1) is an average for all data and time series direction.
  • 1 / 2K is the product of 1 / K, which is the average of K rows, and 1/2, which is the average of the skew-symmetric matrix.
  • 1 / (K-1) is the average of the (K-1) column obtained by subtracting one diagonal column from the K column.
  • N ⁇ (N-1) patterns for example, the following equation (4) can be mentioned as a loss function that considers only the likelihood ratio in which the correct answer class is in the numerator.
  • K is the number of classes
  • M is the number of data
  • T is the time series length.
  • k is a subscript in the row direction
  • l is a subscript in the column direction (that is, a subscript indicating the row number and the column number in the matrix shown in FIG. 6 or the like).
  • is a likelihood ratio, and in the above formula (1), it represents the log-likelihood ratio of the k-th row and l-th column at time t.
  • Logistic is a logistic function and is included as a non-linear function acting on the likelihood ratio ⁇ .
  • equation (4) Since equation (4) considers only the rows of the correct answer class, it includes a step of taking the sum of K rows and averaging the K rows and the skew-symmetric matrix at 1 / 2K as compared with the equation (3) already described. Not done. Further, of the two terms including Kronecker delta in the equation (3), only the previous term remains.
  • the learning unit 300 uses a loss function including a logistical function.
  • a loss function including a logistical function.
  • learning can be performed so that the penalty when the class is incorrect is large and the penalty when the class is correct is small.
  • the sigmoid function ⁇ used in the fifth embodiment changes so that the slope approaches zero (see FIG. 10), while the logistic function changes while the slope is kept constant (see FIG. 11). Therefore, when a loss function including a logistic function is used, the component does not disappear in the process of calculating the slope of the loss function (that is, the process corresponding to the derivative), and more appropriate learning can be performed.
  • the loss functions mentioned in the sixth embodiment are examples, and different loss functions may be created by using a logistic function. Further, the loss function may be created by using another nonlinear function instead of the logistic function. For example, a loss function including a function different from both the sigmoid function and the logistic function described above may be used.
  • the information processing system 1 according to the seventh embodiment will be described with reference to FIGS. 12 and 13.
  • the third embodiment differs from the above-mentioned first to sixth embodiments only in a part of the configuration and operation (specifically, the configuration and operation of the classification device 10), and the other parts are the first. It may be the same as the 1st to 6th embodiments. Therefore, in the following, the parts different from each of the above-described embodiments will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 12 is a block diagram showing a functional configuration of the information processing system according to the seventh embodiment.
  • the same reference numerals are given to the same elements as those of the constituent elements shown in FIG.
  • the likelihood ratio calculation unit 100 in the classification device 10 includes a first calculation unit 110 and a second calculation unit 120.
  • Each of the first calculation unit 110 and the second calculation unit 120 may be realized by, for example, the processor 11 (see FIG. 1) described above.
  • the first calculation unit 110 is configured to be able to calculate the individual likelihood ratio based on two consecutive elements included in the series data.
  • the individual likelihood ratio is calculated as a likelihood ratio indicating the likelihood of a class to which two consecutive elements belong.
  • the first calculation unit 110 may sequentially acquire the elements included in the series data from, for example, the data acquisition unit 50, and sequentially calculate the individual likelihood ratio based on the two consecutive elements.
  • the individual likelihood ratio calculated by the first calculation unit 110 is output to the second calculation unit 120.
  • the second calculation unit 120 is configured to be able to calculate the integrated likelihood ratio based on the plurality of individual likelihood ratios calculated by the first calculation unit 110.
  • the integrated likelihood ratio is calculated as a likelihood ratio indicating the likelihood of a class to which a plurality of elements considered in each of the plurality of individual likelihood ratios belong.
  • the integrated likelihood ratio is calculated as a likelihood ratio indicating the likelihood of a class to which series data including a plurality of elements belong.
  • the integrated likelihood ratio calculated by the second calculation unit 120 is output to the classification unit 200.
  • the class classification unit 200 classifies the series data based on the integrated likelihood ratio.
  • the learning unit 300 may perform learning as a whole of the likelihood ratio calculation unit 100 (that is, the first calculation unit 110 and the second calculation unit 120 together). Learning may be performed separately for 1 calculation unit 110 and 2nd calculation unit 120. Alternatively, the learning unit 300 may be separately provided as a first learning unit that learns only the first calculation unit 110 and a second learning unit that learns only the second calculation unit 120. In this case, only one of the first learning unit and the second learning unit may be provided.
  • FIG. 13 is a flowchart showing the operation flow of the classification device in the information processing system according to the seventh embodiment.
  • the data acquisition unit 50 first acquires the elements included in the series data (step S21).
  • the data acquisition unit 50 outputs the elements of the acquired series data to the first calculation unit 110.
  • the first calculation unit 110 calculates the individual likelihood ratio based on the two consecutive elements acquired (step S22).
  • the second calculation unit 120 calculates the integrated likelihood ratio based on the plurality of individual likelihood ratios calculated by the first calculation unit 110 (step S23).
  • the classification unit 200 performs classification based on the calculated integrated likelihood ratio (step S24).
  • the classification may determine one class to which the series data belongs, or may determine a plurality of classes to which the series data is likely to belong.
  • the class classification unit 200 may output the result of the class classification to a display or the like. Further, the class classification unit 200 may output the result of the class classification by voice via a speaker or the like.
  • an individual likelihood ratio is first calculated based on two elements, and then integrated based on a plurality of individual likelihood ratios.
  • the likelihood ratio is calculated.
  • the class classification can be performed more appropriately by performing the learning by the learning unit 300 described in each of the above-described embodiments. .. That is, by learning so that the penalty when the class is incorrect is large and the penalty when the class is correct is small, more appropriate classification can be performed.
  • the information processing system 1 according to the eighth embodiment will be described with reference to FIGS. 14 and 15.
  • the eighth embodiment differs from the above-mentioned seventh embodiment only in a part of the configuration and operation (specifically, the configuration and operation of the likelihood ratio calculation unit 100), and the other parts are the first. 7 It may be the same as the embodiment. Therefore, in the following, the parts different from each of the above-described embodiments will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 14 is a block diagram showing a functional configuration of the information processing system according to the eighth embodiment.
  • the same elements as those of the constituent elements shown in FIGS. 2 and 12 are designated by the same reference numerals.
  • the likelihood ratio calculation unit 100 in the classification device 10 includes a first calculation unit 110 and a second calculation unit 120.
  • the first calculation unit 110 includes an individual likelihood ratio calculation unit 111 and a first storage unit.
  • the second calculation unit 120 includes an integrated likelihood ratio calculation unit 121 and a second storage unit 122.
  • Each of the individual likelihood ratio calculation unit 111 and the integrated likelihood ratio calculation unit 121 may be realized by, for example, the processor 11 (see FIG. 1) described above.
  • each of the first storage unit 112 and the second storage unit 122 may be realized by, for example, the above-mentioned storage device 14 (see FIG. 1).
  • the individual likelihood ratio calculation unit 111 is configured to be able to calculate the individual likelihood ratio based on two consecutive elements among the elements sequentially acquired by the data acquisition unit 50. More specifically, the individual likelihood ratio calculation unit 111 calculates the individual likelihood ratio based on the newly acquired element and the past data stored in the first storage unit 112. The information stored in the first storage unit 112 is configured to be readable by the individual likelihood ratio calculation unit 111. When the first storage unit 112 stores the past individual likelihood ratio, the individual likelihood ratio calculation unit 111 reads out the stored past individual likelihood ratio and obtains a new individual in consideration of the element. The likelihood ratio may be calculated. On the other hand, when the first storage unit 112 stores the element itself acquired in the past, the individual likelihood ratio calculation unit 111 calculates the past individual likelihood ratio from the stored past elements and newly calculates the individual likelihood ratio. The likelihood ratio to the acquired element may be calculated.
  • the integrated likelihood ratio calculation unit 121 is configured to be able to calculate the integrated likelihood ratio based on a plurality of individual likelihood ratios.
  • the integrated likelihood ratio calculation unit 121 uses the individual likelihood ratio calculated by the individual likelihood ratio calculation unit 111 and the past integrated likelihood ratio stored in the second storage unit 122 to perform a new integrated likelihood ratio. Calculate the likelihood ratio.
  • the information stored in the second storage unit 122 (that is, the past integrated likelihood ratio) is configured to be readable by the integrated likelihood ratio calculation unit 121.
  • FIG. 15 is a flowchart showing the operation flow of the likelihood ratio calculation unit in the information processing system according to the eighth embodiment.
  • the individual likelihood ratio calculation unit 111 in the first calculation unit 110 first collects past data from the first storage unit 112. Read (step S31).
  • the past data is, for example, the processing result (in other words, calculated for the previous element) in the individual likelihood ratio calculation unit 111 of the element acquired immediately before the element acquired this time by the data acquisition unit 50. (Individual likelihood ratio).
  • the past data may be the element itself acquired immediately before the element acquired by acquisition.
  • the individual likelihood ratio calculation unit 111 has a new individual likelihood ratio (that is, a data acquisition unit) based on the elements acquired by the data acquisition unit 50 and the past data read from the first storage unit 112.
  • the individual likelihood ratio to the element acquired this time in 50) is calculated (step S32).
  • the individual likelihood ratio calculation unit 111 outputs the calculated individual likelihood ratio to the second calculation unit 120.
  • the individual likelihood ratio calculation unit 111 may store the calculated individual likelihood ratio in the first storage unit 112.
  • the integrated likelihood ratio calculation unit 121 in the second calculation unit 120 reads out the past integrated likelihood ratio from the second storage unit 122 (step S33).
  • the past integrated likelihood ratio is, for example, the processing result of the integrated likelihood ratio calculation unit 121 (in other words, the previous one) for the element acquired before the element acquired this time by the data acquisition unit 50. It may be the integrated likelihood ratio calculated for the element).
  • the integrated likelihood ratio calculation unit 121 newly integrates based on the likelihood ratio calculated by the individual likelihood ratio calculation unit 111 and the past integrated likelihood ratio read from the second storage unit 122.
  • the likelihood ratio (that is, the integrated likelihood ratio to the element acquired this time by the data acquisition unit 50) is calculated (step S34).
  • the integrated likelihood ratio calculation unit 121 outputs the calculated integrated likelihood ratio to the classification unit 200.
  • the integrated likelihood ratio calculation unit 121 may store the calculated integrated likelihood ratio in the second storage unit 122.
  • the individual likelihood ratio is calculated using the past individual likelihood ratio, and then the past integrated likelihood ratio is used.
  • the integrated likelihood ratio is calculated.
  • the class classification can be more appropriately performed by the learning by the learning unit 300 described in each of the above-described embodiments. It will be possible to do. That is, by learning so that the penalty when the class is incorrect is large and the penalty when the class is correct is small, more appropriate classification can be performed.
  • the information processing system 1 according to the ninth embodiment will be described with reference to FIG.
  • the ninth embodiment differs from the above-mentioned first to seventh embodiments only in a part of the operation (specifically, the operation of the classification unit 200), and the other parts are the first to the first. 7 It may be the same as the embodiment. Therefore, in the following, the parts different from each of the above-described embodiments will be described in detail, and the description of other overlapping parts will be omitted as appropriate.
  • FIG. 16 is a flowchart showing the operation flow of the classification device in the information processing system according to the ninth embodiment.
  • the same reference numerals are given to the same processes as those described with reference to FIG.
  • the data acquisition unit 50 first acquires the elements included in the series data (step S11).
  • the data acquisition unit 50 outputs the elements of the acquired series data to the likelihood ratio calculation unit 100.
  • the likelihood ratio calculation unit 100 calculates the likelihood ratio based on the acquired two or more elements (step S12).
  • the classification unit 200 performs classification based on the calculated likelihood ratio, and in particular in the ninth embodiment, the classification unit 200 selects a plurality of classes to which the series data may belong. And output (step S41). That is, the class classification unit 200 does not determine one class to which the series data belongs, but determines a plurality of classes to which the series data is likely to belong. More specifically, the class classification unit 200 selects k classes (where k is a natural number of n or less) from n classes prepared as classification candidates (where n is a natural number). Execute the process.
  • the class classification unit 200 may output information on k classes to which the series data may belong to a display or the like. Further, the class classification unit 200 may output information about k classes to which the series data may belong by voice via a speaker or the like.
  • the class classification unit 200 When the class classification unit 200 outputs information about k classes to which the series data may belong, they may be rearranged and output. For example, the class classification unit 200 may sort and output information about k classes in descending order of likelihood ratio. Alternatively, the classification unit 200 may output each of the information about the k classes in a different manner for each class. For example, the class classification unit 200 may output a class having a high likelihood ratio in a display mode that emphasizes it, while outputting a class having a low likelihood ratio in a display mode that does not emphasize it. In the case of highlighting, for example, the size and color to be displayed may be changed, or the displayed object may be given motion.
  • the information processing system 1 may be used for proposing a product that the user is likely to be interested in at a shopping site on the Web.
  • the information processing system 1 has k products (that is, k classes) that the user is likely to be interested in from among n products (that is, n classes) that are handled products. May be selected to output to the user (note that k is a number smaller than n).
  • k is a number smaller than n.
  • digital signage it may be possible to capture a user's image by the on-board camera.
  • the user's emotion may be estimated from the user's image, and a store or product corresponding to the emotion may be proposed.
  • the line of sight of the user may be estimated from the image of the user (that is, the portion seen by the user is estimated) to propose a store or product that the user is likely to be interested in.
  • the user's attributes for example, gender, age, etc.
  • weighting may be performed on n classes according to the estimated information.
  • the information processing system 1 according to the ninth embodiment can also be used for criminal investigation. For example, when finding the true criminal among a plurality of suspects, selecting only one of them that is most likely to be the criminal causes a big problem if the selection is wrong. However, in the information processing system 1 according to the present embodiment, it is possible to select and output the top K suspects who are likely to be the criminals. Specifically, the class corresponding to the top k persons who are likely to be the criminals may be selected and output from the series data including the information about each of the plurality of suspects as an element. In this way, for example, it is possible to appropriately find the true criminal by investigating a plurality of suspects who are likely to be criminals.
  • the information processing system 1 according to the ninth embodiment can also be applied to the analysis of radar images. Since many radar images have low sharpness due to their nature, it is difficult to accurately determine what is reflected in the image, for example, only by a machine. However, in the information processing system 1 according to the present embodiment, it is possible to select and output k candidates that are highly likely to be reflected in the radar image. Therefore, it is possible to output k candidates first and have the user judge from among them. For example, if "dog”, “cat”, "ship”, and "tank” are listed as candidates in the radar image of the port, the user can see the "ship” that is highly related to the port in the radar image. You can easily judge that it is in the picture.
  • the above-mentioned application example is an example, and the information processing system 1 according to the present embodiment is applied if it is required to select k from n candidates. Therefore, it is possible to produce a beneficial effect.
  • Each embodiment also implements a processing method in which a program for operating the configuration of the embodiment is recorded on a recording medium so as to realize the functions of the above-described embodiments, the program recorded on the recording medium is read out as a code, and the program is executed by a computer. Included in the category of morphology. That is, a computer-readable recording medium is also included in the scope of each embodiment. Further, not only the recording medium on which the above-mentioned program is recorded but also the program itself is included in each embodiment.
  • the recording medium for example, a floppy (registered trademark) disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a magnetic tape, a non-volatile memory card, or a ROM can be used.
  • a floppy (registered trademark) disk for example, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a magnetic tape, a non-volatile memory card, or a ROM
  • the program recorded on the recording medium is executed by itself, but also the program that operates on the OS and executes the process in cooperation with other software and the function of the expansion board is also an embodiment. Is included in the category of.
  • the information processing system is based on an acquisition means for acquiring a plurality of elements included in the series data and at least two consecutive elements among the plurality of elements, and the likelihood of the class to which the series data belongs.
  • a calculation means for calculating the likelihood ratio indicating the above a classification means for classifying the series data into at least one class among a plurality of classes that are classification candidates based on the likelihood ratio, and a correct answer to which the series data belongs.
  • the likelihood is increased when the class is in the molecule of the likelihood ratio, and the likelihood ratio is decreased when the correct class is in the denominator of the likelihood ratio.
  • It is an information processing system characterized by being provided with a learning means for learning about calculation of a ratio.
  • the learning means has a denominator of the likelihood of belonging to one of the N classes (where N is a natural number) that are candidates for classification of the series data.
  • the present invention is characterized in that the learning is performed using the loss function considering the likelihood ratio of the N ⁇ (N-1) pattern having the likelihood indicating the likelihood of belonging to another class as a molecule. It is an information processing system described in.
  • Appendix 3 The information processing system according to Appendix 3 is characterized in that the learning means performs the learning by using a loss function in consideration of the likelihood ratio of a part of the N ⁇ (N-1) patterns. This is the information processing system described in Appendix 2.
  • Appendix 4 In the information processing system described in Appendix 4, the learning means performs the learning by using a loss function in consideration of the likelihood ratio in which the correct answer class is in the molecule among the N ⁇ (N-1) patterns.
  • the information processing system according to Supplementary Note 5 is the information processing according to any one of Supplementary note 1 to 4, wherein the loss function includes a sigmoid function as a nonlinear function acting on the likelihood ratio. It is a system.
  • the information processing system according to Supplementary Note 6 is the information processing according to any one of Supplementary note 1 to 4, wherein the loss function includes a logistic function as a nonlinear function acting on the likelihood ratio. It is a system.
  • the likelihood ratio is an integrated likelihood ratio calculated by considering a plurality of individual likelihood ratios calculated based on two consecutive elements included in the series data.
  • the information processing system according to any one of Supplementary note 1 to 6, characterized in that it is present.
  • Appendix 8 In the information processing system according to Appendix 8, the acquisition means sequentially acquires a plurality of elements included in the series data, and the calculation means calculates the individual likelihood ratio based on the newly acquired elements.
  • the information processing method acquires a plurality of elements included in the series data, and shows the likelihood of the class to which the series data belongs based on at least two consecutive elements among the plurality of elements.
  • the degree ratio is calculated, and the series data is classified into at least one class among a plurality of classes that are candidates for classification based on the likelihood ratio, and the correct answer class to which the series data belongs becomes the molecule of the likelihood ratio.
  • Learning about the calculation of the likelihood ratio is performed using a loss function in which the likelihood ratio becomes large in a certain case and the likelihood ratio becomes small when the correct answer class is in the denominator of the likelihood ratio. This is a characteristic information processing method.
  • Appendix 10 The computer program according to Appendix 10 acquires a plurality of elements included in the series data, and based on at least two consecutive elements among the plurality of elements, a likelihood indicating the likelihood of the class to which the series data belongs. The ratio is calculated, the series data is classified into at least one class among a plurality of classes that are candidates for classification based on the likelihood ratio, and the correct answer class to which the series data belongs is in the molecule of the likelihood ratio.
  • a computer is used to learn about the calculation of the likelihood ratio by using a loss function in which the likelihood ratio is increased and the correct answer class is in the denominator of the likelihood ratio and the likelihood ratio is decreased. It is a computer program characterized by operating.
  • Appendix 11 The recording medium described in Appendix 11 is a recording medium characterized in that the computer program described in Appendix 10 is recorded.
  • Information processing system 11 Processor 14 Storage device 10
  • Classification device 50 Data acquisition unit 100 Probability ratio calculation unit 110 First calculation unit 111 Individual likelihood ratio calculation unit 112 First storage unit 120 Second calculation unit 121 Integrated likelihood ratio calculation Part 122 2nd storage part 200 Class classification part 300 Learning part

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Abstract

情報処理システム(1)は、系列データに含まれる複数の要素を取得する取得手段(50)と、複数の要素のうち少なくとも2つの連続する要素に基づいて、系列データが属するクラスの尤もらしさを示す尤度比を算出する算出手段(100)と、尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに系列データを分類する分類手段(200)と、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる損失関数を用いて、尤度比の算出に関する学習を行う学習手段(300)とを備える。このような情報処理システムによれば、分類候補である複数のクラスの中から、系列データが属するクラス適切に選択することが可能である。

Description

情報処理システム、情報処理方法、及びコンピュータプログラム
 この開示は、例えばクラス分類に関する情報を処理する情報処理システム、情報処理方法、及びコンピュータプログラムの技術分野に関する。
 この種のシステムとして、尤度に関する学習処理を実行するものが知られている。例えば特許文献1では、尤度の判定に用いるサポートベクターマシンを学習することが開示されている。特許文献2では、尤度を利用する識別機の学習に、サポートベクターマシンやロジスティック回帰を用いてもよいことが開示されている。
 その他の関連する技術として、例えば特許文献3では、画像に含まれる人物が登録人物か否かを判定する装置において、所定の認識条件及び優勢条件に基づいて画像ファイルに対するフォルダ分類処理を行う技術が開示されている。
特開2019-139618号公報 特開2017-174054号公報 特開2010-250730号公報
 この開示は、上述した関連する技術を改善することを目的とする。
 この開示の情報処理システムの一の態様は、系列データに含まれる複数の要素を取得する取得手段と、前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出する算出手段と、前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類する分類手段と、前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行う学習手段とを備える。
 この開示の情報処理方法の一の態様は、系列データに含まれる複数の要素を取得し、前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出し、前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類し、前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行う。
 この開示のコンピュータプログラムの一の態様は、系列データに含まれる複数の要素を取得し、前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出し、前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類し、前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行うようにコンピュータを動作させる。
第1実施形態に係る情報処理システムのハードウェア構成を示すブロック図である。 第1実施形態に係る情報処理システムの機能的構成を示すブロック図である。 第1実施形態に係る情報処理システムにおける分類装置の動作の流れを示すフローチャートである。 第1実施形態に係る情報処理システムにおける学習部の動作の流れを示すフローチャートである。 第2実施形態に係る情報処理システムにおける学習部の動作の流れを示すフローチャートである。 第2実施形態に係る情報処理システムにおける学習部が考慮する尤度比の一例を示すマトリクス図である。 第3実施形態に係る情報処理システムにおける学習部の動作の流れを示すフローチャートである。 第4実施形態に係る情報処理システムにおける学習部の動作の流れを示すフローチャートである。 第4実施形態に係る情報処理システムにおける学習部が考慮する尤度比の一例を示すマトリクス図である。 第5実施形態に係る情報処理システムで用いられるシグモイド関数の一例を示すグラフである。 第6実施形態に係る情報処理システムで用いられるロジスティック関数の一例を示すグラフである。 第7実施形態に係る情報処理システムの機能的構成を示すブロック図である。 第7実施形態に係る情報処理システムにおける分類装置の動作の流れを示すフローチャートである。 第8実施形態に係る情報処理システムの機能的構成を示すブロック図である。 第8実施形態に係る情報処理システムにおける尤度比算出部の動作の流れを示すフローチャートである。 第9実施形態に係る情報処理システムにおける分類装置の動作の流れを示すフローチャートである。
 以下、図面を参照しながら、情報処理システム、情報処理方法、及びコンピュータプログラムの実施形態について説明する。
 <第1実施形態>
 第1実施形態に係る情報処理システムについて、図1から図4を参照して説明する。
 (ハードウェア構成)
 まず、図1を参照しながら、第1実施形態に係る情報処理システムのハードウェア構成について説明する。図1は、第1実施形態に係る情報処理システムのハードウェア構成を示すブロック図である。
 図1に示すように、第1実施形態に係る情報処理システム1は、プロセッサ11と、RAM(Random Access Memory)12と、ROM(Read Only Memory)13と、記憶装置14とを備えている。情報処理システム1は更に、入力装置15と、出力装置16とを備えていてもよい。プロセッサ11と、RAM12と、ROM13と、記憶装置14と、入力装置15と、出力装置16とは、データバス17を介して接続されている。
 プロセッサ11は、コンピュータプログラムを読み込む。例えば、プロセッサ11は、RAM12、ROM13及び記憶装置14のうちの少なくとも一つが記憶しているコンピュータプログラムを読み込むように構成されている。或いは、プロセッサ11は、コンピュータで読み取り可能な記録媒体が記憶しているコンピュータプログラムを、図示しない記録媒体読み取り装置を用いて読み込んでもよい。プロセッサ11は、ネットワークインタフェースを介して、情報処理システム1の外部に配置される不図示の装置からコンピュータプログラムを取得してもよい(つまり、読み込んでもよい)。プロセッサ11は、読み込んだコンピュータプログラムを実行することで、RAM12、記憶装置14、入力装置15及び出力装置16を制御する。本実施形態では特に、プロセッサ11が読み込んだコンピュータプログラムを実行すると、プロセッサ11内には、尤度比を用いる分類及びそれに関する学習処理を実行ための機能ブロックが実現される。なお、プロセッサ11の一例として、CPU(Central Processing Unit)、GPU(Graphics Processing Unit)、FPGA(field-programmable gate array)、DSP(Demand-Side Platform)、ASIC(Application Specific Integrated Circuit)が挙げられる。プロセッサ11は、上述した一例のうち一つを用いてもよいし、複数を並列で用いてもよい。
 RAM12は、プロセッサ11が実行するコンピュータプログラムを一時的に記憶する。RAM12は、プロセッサ11がコンピュータプログラムを実行している際にプロセッサ11が一時的に使用するデータを一時的に記憶する。RAM12は、例えば、D-RAM(Dynamic RAM)であってもよい。
 ROM13は、プロセッサ11が実行するコンピュータプログラムを記憶する。ROM13は、その他に固定的なデータを記憶していてもよい。ROM13は、例えば、P-ROM(Programmable ROM)であってもよい。
 記憶装置14は、情報処理システム1が長期的に保存するデータを記憶する。記憶装置14は、プロセッサ11の一時記憶装置として動作してもよい。記憶装置14は、例えば、ハードディスク装置、光磁気ディスク装置、SSD(Solid State Drive)及びディスクアレイ装置のうちの少なくとも一つを含んでいてもよい。
 入力装置15は、情報処理システム1のユーザからの入力指示を受け取る装置である。入力装置15は、例えば、キーボード、マウス及びタッチパネルのうちの少なくとも一つを含んでいてもよい。入力装置15は、専用のコントローラ(操作端末)であってもよい。また、入力装置15は、ユーザが保有する端末(例えば、スマートフォンやタブレット端末等)を含んでいてもよい。入力装置15は、例えばマイクを含む音声入力が可能な装置であってもよい。
 出力装置16は、情報処理システム1に関する情報を外部に対して出力する装置である。例えば、出力装置16は、情報処理システム1に関する情報を表示可能な表示装置(例えば、ディスプレイ)であってもよい。ここでの表示装置は、テレビモニタ、パソコンモニタ、スマートフォンのモニタ、タブレット端末のモニタ、その他の携帯端末のモニタであってよい。また、表示装置は、店舗等の各種施設に設置される大型モニタやデジタルサイネージ等であってよい。また、出力装置16は、画像以外の形式で情報を出力する装置であってもよい。例えば、出力装置16は、情報処理システム1に関する情報を音声で出力するスピーカであってもよい。
 (機能的構成)
 次に、図2を参照しながら、第1実施形態に係る情報処理システム1の機能的構成について説明する。図2は、第1実施形態に係る情報処理システムの機能的構成を示すブロック図である。
 図2に示すように、第1実施形態に係る情報処理システム1は、分類装置10と、学習部300と備えている。分類装置10は、入力される系列データのクラス分類を行う装置であり、その機能を実現するための処理ブロックとして、データ取得部50と、尤度比算出部100と、クラス分類部200とを備えて構成されている。また、学習部300は、分類装置10に関する学習処理を実行可能に構成されている。なお、ここでは、分類装置10とは別に学習部300が設けられる例を挙げているが、分類装置10が学習部300を備えて構成されてもよい。データ取得部50、尤度比算出部100、クラス分類部200、及び学習部300の各々は、上述したプロセッサ11(図1参照)によって実現されてよい。
 データ取得部50は、系列データに含まれる複数の要素を取得可能に構成されている。データ取得部50は、任意のデータ取得装置(例えば、カメラやマイク等)から直接データを取得するものであってもよいし、あらかじめデータ取得装置で取得されストレージ等に記憶されているデータを読み出すものであってもよい。カメラからデータを取得する場合、データ取得部50は複数のカメラの各々からデータを取得するように構成されていてもよい。データ取得部50で取得された系列データの要素は、尤度比算出部100に出力される構成となっている。なお、系列データとは、所定の順番で並んだ複数の要素を含むデータであり、例えば時系列データが一例として挙げられる。系列データのより具体的な例としては、動画データや音声データが挙げられるが、これに限られるものではない。
 尤度比算出部100は、データ取得部50で取得された複数の要素のうち少なくとも2つの連続する要素に基づいて、尤度比を算出可能に構成されている。なお、ここでの「尤度比」とは、系列データが属するクラスの尤もらしさを示す指標である。尤度比の具体例や具体的な算出方法については、後述する他の実施形態において詳しく説明する。
 クラス分類部200は、尤度比算出部100で算出された尤度比に基づいて、系列データを分類可能に構成されている。クラス分類部200は、分類候補である複数のクラスの中から、系列データが属する少なくとも1つのクラスを選択する。分類候補である複数のクラスは、予め設定されたものであってもよい。或いは、分類候補である複数のクラスは、ユーザによって適宜設定されるものであってもよいし、扱う系列データの種別等に基づいて適宜設定されるものであってもよい。
 学習部300は、損失関数を用いて尤度比の算出に関する学習を行う。具体的には、尤度比に基づくクラス分類が正確に行われるように、尤度比の算出に関する学習を実行する。学習部300が用いる損失関数は、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる関数として定義されるものである。損失関数は、このような定義を満たす関数として予め設定されていてよい。損失関数の具体例については、後述する他の実施形態において詳しく説明する。
 (分類動作の流れ)
 次に、図3を参照しながら、第1実施形態に係る情報処理システム1における分類装置10の動作(具体的には、学習後のクラス分類動作)の流れについて説明する。図3は、第1実施形態に係る情報処理システムにおける分類装置の動作の流れを示すフローチャートである。
 図3に示すように、分類装置10の動作が開始されると、まずデータ取得部50が、系列データに含まれる要素を取得する(ステップS11)。データ取得部50は、取得した系列データの要素を、尤度比算出部100に出力する。そして、尤度比算出部100は、取得された2つ以上の要素に基づいて尤度比を算出する(ステップS12)。
 続いて、クラス分類部200が、算出された尤度比に基づいてクラス分類を行う(ステップS13)。クラス分類は、系列データが属する1つのクラスを決定するものであってもよいし、系列データが属する可能性の高い複数のクラスを決定するものであってもよい。クラス分類部200は、クラス分類の結果をディスプレイ等に出力するようにしてもよい。また、クラス分類部200は、クラス分類の結果を、スピーカ等を介して音声で出力するようにしてもよい。
 (学習動作の流れ)
 次に、図4を参照しながら、第1実施形態に係る情報処理システム1における学習部300の動作(即ち、尤度比の算出に関する学習動作)の流れについて説明する。図4は、第1実施形態に係る情報処理システムにおける学習部の動作の流れを示すフローチャートである。
 図4に示すように、学習動作が開始されると、まず学習部300には訓練データが入力される(ステップS101)。訓練データは、例えば系列データと、その系列データが属する正解クラスに関する情報(即ち、正解データ)とのセットとして構成されてよい。
 続いて、学習部300は、入力された訓練データを用いて損失関数を算出する(ステップS102)。ここでの損失関数は、すでに説明したように、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる関数である。
 続いて、学習部300は、算出した損失関数が小さくなるようにパラメータ(具体的には、尤度比を算出するモデルのパラメータ)を調整する(ステップS103)。即ち、学習部300は、尤度比を算出するモデルのパラメータを最適化する。なお、損失関数を用いたパラメータの最適化手法としては、適宜既存の技術を採用することが可能である。最適化手法の一例としては、誤差逆伝播法が挙げられるが、その他の手法を用いてもよい。
 その後、学習部300は、すべての学習が終了したか否かを判定する(ステップS104)。学習部300は、例えばすべての訓練データが入力されたか否かによって、すべての学習が終了したか否かを判定してよい。或いは、学習部300は、学習開始から所定期間が経過したか否かによって、すべての学習が終了したか否かを判定してよい。或いは、学習部300は、上述したステップS101からS103の処理が所定回数ループされたか否かによって、すべての学習が終了したか否かを判定してよい。
 すべての学習が終了したと判定した場合(ステップS104:YES)、一連の処理は終了する。一方で、すべての学習が終了していないと判定した場合(ステップS104:NO)、学習部300は、再びステップS101から処理を開始する。これにより、訓練データを用いる学習処理が繰り返され、パラメータがより最適なものへと調整されていく。
 (技術的効果)
 次に、第1実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図1から図4で説明したように、第1実施形態に係る情報処理システム1では、学習部300によって、クラス分類に用いる尤度比の算出に関する学習が実行される。そして本実施形態では特に、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる損失関数を用いて学習が行われる。このような損失関数を用いれば、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習が行える。この結果、分類候補である複数のクラスの中から、系列データが属する少なくとも1つのクラスを適切に選択することが可能となる。
 分類候補として複数のクラスが存在する場合(所謂、マルチクラス分類を行う場合)、学習時にどのような尤度比を考慮するのか(例えば、何と何の比をとればよいのか)を決定することが容易ではない。しかるに、上述した損失関数を利用すれば、正解クラスが尤度比の分子にあるのか、それとも分母にあるのかによって尤度比の大小が変化し、損失関数に対して与える影響が変化する。このような損失関数を用いれば、マルチクラス分類における尤度比の算出に関する学習を適切に実行することができる。その結果、適切なクラス分類を実現することが可能となる。なお、学習時にどのような尤度比を考慮するのかについては、分類候補として3つ以上のクラスが存在する場合に、特に決定するのが難しくなる。よって、本実施形態に係る技術的効果は、分類候補が3つ以上のクラスである場合に顕著に発揮される。
 <第2実施形態>
 第2実施形態に係る情報処理システム1について、図5及び図6を参照して説明する。なお、第2実施形態は、上述した第1実施形態と比較して一部の動作が異なるのみであり、例えば装置構成(図1及び図2参照)や、分類装置10の動作(図3参照)等については、第1実施形態と同様であってよい。このため、以下では、第1実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (学習動作の流れ)
 まず、図5を参照しながら、第2実施形態に係る情報処理システム1における学習部300の動作の流れについて説明する。図5は、第2実施形態に係る情報処理システムの動作の流れを示すフローチャートである。なお、図5では、図4で示した処理と同様の処理に同一の符号を付している。
 図5に示すように、第2実施形態に係る情報処理システム1の動作が開始されると、まず学習部300には訓練データが入力される(ステップS101)。
 続いて、学習部300は、入力された訓練データを用いて損失関数を算出するが、第2実施形態では特に、系列データの分類候補であるN個(ただし、Nは自然数)のクラスのうち、一のクラスに属する尤もらしさを示す尤度を分母とし、他のクラスに属する尤もらしさを示す尤度を分子とするN×(N-1)パターンの尤度比を考慮した損失関数を算出する(ステップS201)。この損失関数も、第1実施形態と同様に、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる関数である。損失関数に考慮される尤度比については、後に具体例を挙げて詳しく説明する。
 続いて、学習部300は、算出した損失関数が小さくなるようにパラメータを調整する(ステップS103)。即ち、学習部300は、尤度比を算出するモデルのパラメータを最適化する。その後、学習部300は、すべての学習が終了したか否かを判定する(ステップS104)。すべての学習が終了したと判定した場合(ステップS104:YES)、一連の処理は終了する。一方で、すべての学習が終了していないと判定した場合(ステップS104:NO)、学習部300は、再びステップS101から処理を開始する。
 (考慮する尤度比の具体例)
 次に、図6を参照しながら、上述した学習部300による学習動作において考慮される尤度比(即ち、損失関数の算出に考慮される尤度比)について、具体的に説明する。図6は、第2実施形態に係る情報処理システムにおける学習部が考慮する尤度比の一例を示すマトリクス図である。
 図6に示すように、尤度をマトリクス状にして考える。ここでは説明の便宜上、分類候補であるクラスが、「クラス0」、「クラス1」、「クラス2」の3つであるとする。p(X|y=0)は、系列データが「クラス0」である尤もらしさを示す尤度である。p(X|y=1)は、系列データが「クラス1」である尤もらしさを示す尤度である。p(X|y=2)は、系列データが「クラス2」である尤もらしさを示す尤度である。
 マトリクスの上から1行目は、対数尤度比(以下、単に「尤度比」と称する)の分子が、すべてp(X|y=0)となっている。マトリクスの上から2行目は、尤度比の分子が、すべてp(X|y=1)となっている。マトリクスの上から3行目は、尤度比の分子が、すべてp(X|y=2)となっている。一方、マトリクスの左から1列目は、尤度比の分母が、すべてp(X|y=0)となっている。マトリクスの左から2列目は、尤度比の分母が、すべてp(X|y=1)となっている。マトリクスの左から3列目は、尤度比の分母が、すべてp(X|y=2)となっている。
 マトリクスの対角線上にある尤度比(図6においてグレーで網掛けされている尤度比)は、分母と分子とが互いに同じ尤度になっている。具体的には、上から1行目、左から1列目のlog{p(X|y=0)/p(X|y=0)}、上から2行目、左から2列目のlog{p(X|y=1)/p(X|y=1)}、上から3行目、左から3列目のlog{p(X|y=2)/p(X|y=2)}は、それぞれ分母と分子とが同じである。また、上記対角線上の尤度比を挟んで対向する位置にある尤度比は、互いに分母と分子とが逆になっている。具体的には、上から1行目、左から2列目のlog{p(X|y=0)/p(X|y=1)}と、上から2行目、左から1列目のlog{p(X|y=1)/p(X|y=0)}とでは、分母と分子が逆になっている。同様に、上から1行目、左から3列目のlog{p(X|y=0)/p(X|y=2)}と、上から3行目、左から1列目のlog{p(X|y=2)/p(X|y=0)}とでは、分母と分子が逆になっている。上から2行目、左から3列目のlog{p(X|y=1)/p(X|y=2)}と、上から3行目、左から2列目のlog{p(X|y=2)/p(X|y=1)}とでは、分母と分子が逆になっている。よって、対角線を挟んで対向する位置にある尤度比は、互いに符号が逆の値となる。このように、マトリクスで示される各尤度比は、交代行列のように並んでいる。
 ここで特に、対角線上にある分母と分子とが同じになる尤度比は、すべてlog1であり、その値はゼロである。このため、対角線上にある分母と分子が同じになる尤度比については、損失関数に考慮しても実質的に意味がない値となる。このため、対角線上にある分母と分子とが同じになる尤度比については、損失関数に考慮しないものとする。対角線上の尤度比を除いた残りの尤度比の数は、クラスの数をNとすると、N×(N-1)となる。本実施形態では、これらのN×(N-1)パターンの尤度比(即ち、マトリクスにおける対角線上の尤度比を除く尤度比)を損失関数に考慮する。なお、N×(N-1)パターンの尤度比を考慮した損失関数の具体例については、後述する他の実施形態において詳しく説明する。
 (技術的効果)
 次に、第2実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図5及び図6で説明したように、第2実施形態に係る情報処理システム1では、一のクラスに属する尤もらしさを示す尤度を分母とし、他のクラスに属する尤もらしさを示す尤度を分子とするN×(N-1)パターンの尤度比を考慮した損失関数を用いて学習が実行される。このような損失関数を用いれば、第1実施形態と同様に、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習が行える。この結果、分類候補である複数のクラスの中から、系列データが属する少なくとも1つのクラスを適切に選択することが可能となる。
 <第3実施形態>
 第3実施形態に係る情報処理システム1について、図7を参照して説明する。なお、第3実施形態は、上述した第1及び第2実施形態と一部の動作が異なるのみで、その他の部分については、第1及び第2実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (学習動作の流れ)
 まず、図7を参照しながら、第3実施形態に係る情報処理システム1における学習部300の動作の流れについて説明する。図7は、第3実施形態に係る情報処理システムの動作の流れを示すフローチャートである。なお、図7では、図4で示した処理と同様の処理に同一の符号を付している。
 図7に示すように、第3実施形態に係る情報処理システム1の動作が開始されると、まず学習部300には訓練データが入力される(ステップS101)。
 続いて、学習部300は、入力された訓練データを用いて損失関数を算出するが、第3実施形態では特に、系列データの分類候補であるN個のクラスのうち、一のクラスに属する尤もらしさを示す尤度を分母とし、他のクラスに属する尤もらしさを示す尤度を分子とするN×(N-1)パターンの尤度比の一部を考慮した損失関数を算出する(ステップS301)。即ち、第3実施形態に係る学習部300は、第2実施形態で説明したN×(N-1)パターンの尤度比をすべて考慮するのではなく、その一部のみを考慮する。なお、この損失関数も、第1実施形態と同様に、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる関数である。N×(N-1)パターンの一部の尤度比を考慮した損失関数の具体例については、後述する他の実施形態において詳しく説明する。
 続いて、学習部300は、算出した損失関数が小さくなるようにパラメータを調整する(ステップS103)。その後、学習部300は、すべての学習が終了したか否かを判定する(ステップS104)。すべての学習が終了したと判定した場合(ステップS104:YES)、一連の処理は終了する。一方で、すべての学習が終了していないと判定した場合(ステップS104:NO)、学習部300は、再びステップS101から処理を開始する。
 (考慮する尤度比の選択例)
 続いて、損失関数に考慮する尤度比の選択例(即ち、N×(N-1)パターンの一部の尤度比の選択例)について具体的に説明する。
 N×(N-1)パターンの尤度比のうち、損失関数に考慮する一部の尤度比については、予めユーザ等が選択してもよいし、学習部300が自動的に選択するようにしてもよい。学習部300が損失関数に考慮する一部の尤度比を選択する場合、学習部300は、予め設定された所定のルールに従って尤度比を選択してもよい。或いは、学習部300は、算出された尤度比の値に基づいて選択するか否かを決定してもよい。
 損失関数に考慮する一部の尤度比の選択例としては、例えば図6に示したマトリクスにおける1つの行又は1つの列の尤度比のみを選択するものが挙げられる。例えば、損失関数に考慮する尤度比として、図6に示すマトリクスの1行目の尤度比だけを選択してもよいし、2行目の尤度比だけを選択してもよいし、3行目の尤度比だけを選択してもよい。或いは、マトリクスの1列目の尤度比だけを選択してもよいし、2列目の尤度比だけを選択してもよいし、3列目の尤度比だけを選択してもよい。
 また、マトリクスにおける一部の複数行又は一部の複数列の尤度比のみを選択してもよい。具体的には、マトリクスの1行目及び2行目の尤度比だけを選択してもよいし、2行目及び3行目の尤度比だけを選択してもよいし、3行目及び1行目の尤度比だけを選択してもよい。或いは、マトリクスの1列目及び2列目の尤度比だけを選択してもよいし、2列目及び3列目の尤度比だけを選択してもよいし、3列目及び1列目の尤度比だけを選択してもよい。
 なお、上述した尤度比の選択例は、あくまで一例であり、その他の尤度比が損失関数に考慮される尤度比として選択されてもよい。例えば、損失関数に考慮される尤度比は、行や列に関係なくランダムに選択されてもよい。
 (技術的効果)
 次に、第3実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図7で説明したように、第3実施形態に係る情報処理システム1では、一のクラスに属する尤もらしさを示す尤度を分母とし、他のクラスに属する尤もらしさを示す尤度を分子とするN×(N-1)パターンのうち、一部の尤度比を考慮した損失関数を用いて学習が実行される。このような損失関数を用いれば、上述した各実施形態と同様に、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習が行える。この結果、分類候補である複数のクラスの中から、系列データが属する少なくとも1つのクラスを適切に選択することが可能となる。また、第3実施形態では特に、損失関数に考慮する尤度比をN×(N-1)パターンの中から適切に選択することで、N×(N-1)パターンのすべての尤度比を考慮する場合と比べて、効率的な学習を行うことができる。例えば、損失関数に与える影響が比較的大きい尤度比のみを選択し、損失関数に与える影響が比較的小さい尤度比を選択しないようにすれば、学習効率を高めることができる。
 <第4実施形態>
 第4実施形態に係る情報処理システム1について、図8及び図9を参照して説明する。なお、第4実施形態は、上述した第3実施形態の具体的な選択例(即ち、損失関数に考慮する一部の尤度比の選択例)を説明するものであり、その他の部分については、第3実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (学習動作の流れ)
 まず、図8を参照しながら、第4実施形態に係る情報処理システム1における学習部300の動作の流れについて説明する。図8は、第4実施形態に係る情報処理システムの動作の流れを示すフローチャートである。なお、図8では、図4で示した処理と同様の処理に同一の符号を付している。
 図8に示すように、第4実施形態に係る情報処理システム1の動作が開始されると、まず学習部300には訓練データが入力される(ステップS101)。
 続いて、学習部300は、入力された訓練データを用いて損失関数を算出するが、第4実施形態では特に、すでに説明したN×(N-1)パターンの尤度比のうち、正解クラスが分子にある尤度比を考慮した損失関数を算出する(ステップS401)。即ち、第4実施形態に係る学習部300は、第3実施形態で説明したN×(N-1)パターンの一部の尤度比として、正解クラスが分子にある尤度比を選択する。なお、この損失関数も、第1実施形態と同様に、系列データが属する正解クラスが尤度比の分子にある場合に尤度比が大きくなり、正解クラスが尤度比の分母にある場合に尤度比が小さくなる関数である。正解クラスが分子にある尤度比を考慮した損失関数の具体例については、後述する他の実施形態において詳しく説明する。
 続いて、学習部300は、算出した損失関数が小さくなるようにパラメータを調整する(ステップS103)。その後、学習部300は、すべての学習が終了したか否かを判定する(ステップS104)。すべての学習が終了したと判定した場合(ステップS104:YES)、一連の処理は終了する。一方で、すべての学習が終了していないと判定した場合(ステップS104:NO)、学習部300は、再びステップS101から処理を開始する。
 (考慮する尤度比の具体例)
 次に、図9を参照しながら、上述した学習部300による学習動作において考慮される尤度比(即ち、損失関数の算出に考慮される尤度比)について、具体的に説明する。図9は、第4実施形態に係る情報処理システムにおける学習部が考慮する尤度比の一例を示すマトリクス図である。
 図9に示すマトリクスは、すでに第2実施形態で説明したように(図6参照)、各尤度比が交代行列のように並んでいる。第4実施形態に係る学習部300は、このようなマトリクスにおける対角線上の尤度比を除いたN×(N-1)パターンの尤度比の中から、正解クラスが分子にある尤度比を選択して損失関数に考慮する。
 例えば、訓練データとして入力された系列データの正解クラスが「クラス1」であるとする。この場合、学習部300は、N×(N-1)パターンの尤度比の中から、クラス1が分子にある尤度比を選択して損失関数に考慮する。具体的には、図9の上から2行目の尤度比(対角線上の尤度比を除く)のみを選択して、損失関数に考慮する。この場合、上から2行目、左から1列目のlog{p(X|y=1)/p(X|y=0)}と、上から2行目、左から3列目のlog{p(X|y=1)/p(X|y=2)}とが損失関数に考慮されることになる。即ち、図9中のグレーで網掛けされていない尤度比が損失関数に考慮されることになる。
 なお、訓練データとして入力された系列データの正解クラスが「クラス0」である場合、学習部300は、N×(N-1)パターンの尤度比の中から、クラス0が分子にある尤度比を選択して損失関数に考慮すればよい。具体的には、図9の上から1行目の尤度比(対角線上の尤度比を除く)のみを選択して、損失関数に考慮すればよい。この場合、上から1行目、左から2列目のlog{p(X|y=0)/p(X|y=1)}と、上から1行目、左から3列目のlog{p(X|y=0)/p(X|y=2)}とが損失関数に考慮されることになる。
 同様に、訓練データとして入力された系列データの正解クラスが「クラス2」である場合、学習部300は、N×(N-1)パターンの尤度比の中から、クラス2が分子にある尤度比を選択して損失関数に考慮すればよい。具体的には、図9の上から3行目の尤度比(対角線上の尤度比を除く)のみを選択して、損失関数に考慮すればよい。この場合、上から3行目、左から1列目のlog{p(X|y=2)/p(X|y=0)}と、上から3行目、左から2列目のlog{p(X|y=2)/p(X|y=1)}とが損失関数に考慮されることになる。
 (技術的効果)
 次に、第4実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図8及び図9で説明したように、第4実施形態に係る情報処理システム1では、N×(N-1)パターンのうち、正解クラスが分子にある尤度比を考慮した損失関数を用いて学習が実行される。このような損失関数を用いれば、上述した各実施形態と同様に、適切な学習が行われることになるため、分類候補である複数のクラスの中から、系列データが属する少なくとも1つのクラスを適切に選択することが可能となる。また、第4実施形態では特に、正解クラスが分子にある尤度比(言い換えれば、損失関数に大きな影響を与える可能性がある尤度比)が損失関数に考慮されるため、N×(N-1)パターンのすべての尤度比を考慮する場合と比べて、効率的な学習を行うことができる。
 <第5実施形態>
 第5実施形態に係る情報処理システム1について、図10を参照して説明する。なお、第5実施形態は、上述した第1から第4実施形態で用いられる損失関数の具体例を説明するものであり、装置構成や動作の流れについては、第1から第4実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (シグモイド関数を含む損失関数)
 まず、図10を参照しながら、第5実施形態に係る情報処理システム1で用いられる損失関数の概要について説明する。図10は、第5実施形態に係る情報処理システムで用いられるシグモイド関数の一例を示すグラフである。
 図10に示すように、第5実施形態に係る情報処理システム1では、学習部300がシグモイド関数を含む損失関数を用いて学習を行う。より具体的には、学習部300は、シグモイド関数を、尤度比に作用する非線形関数として含む損失関数を用いて学習を行う。以下では、シグモイド関数を含む損失関数の具体例を挙げて説明を進める。
 (すべての尤度比を考慮する例)
 1つ目の例として、第2実施形態のように、N×(N-1)パターンのすべての尤度比を考慮する場合の損失関数を挙げて説明する。なお、以下で示す式の尤度比は、対数尤度比(LLR:Log Likehood Ratio)であるものとする。
 N×(N-1)パターンのすべての尤度比を考慮する損失関数として、例えば下記式(1)が挙げられる。
Figure JPOXMLDOC01-appb-M000001
 上記式(1)において、Kはクラス数、Mはデータ数、Tは時系列長である。また、kは行方向、lは列方向の添字(即ち、図6等に示すマトリクスにおける行番号及び列番号を示す添字)である。δはクロネッカーのデルタであり、添字が一致する場合に「1」、それ以外の場合は「0」となる。λは尤度比であり、上記数式(1)では、時刻tにおけるk行l列目の対数尤度比を表している。σは、シグモイド関数であり、尤度比λに作用する非線形関数として含まれている。
 式(1)における1/MTは、データ及び時系列方向の全てに対し平均をとるものである。1/2Kは、K行の平均をとる1/Kと、交代行列の平均をとる1/2とを乗じたものである。1/(K-1)は、K列から対角線の1列を差し引いた(K-1)列の平均をとるものである。
 式(1)の損失関数Lでは、k行が正解クラスであれば、クロネッカーのδが「1」となり、k行が不正解クラス(即ち、正解クラス以外のクラス)であれば、クロネッカーのδが「0」となる。この結果、k行が正解クラスであれば(言い換えれば、正解クラスが尤度比の分子にあれば)、尤度比の値が大きくなる。一方で、k行が不正解クラスであれば(言い換えれば、不正解クラスが尤度比の分子にあれば)、尤度比の値が小さくなる。
 (一部の尤度比を考慮する例)
 次に、2つ目の例として、第3及び第4実施形態のように、N×(N-1)パターンのうち一部の尤度比を考慮する場合の損失関数を挙げて説明する。以下では特に、第4実施形態のように、N×(N-1)パターンのうち、正解クラスが分子にある尤度比のみを考慮する例を挙げて説明する。
 N×(N-1)パターンのうち、正解クラスが分子にある尤度比のみを考慮する損失関数として、例えば下記式(2)が挙げられる。
Figure JPOXMLDOC01-appb-M000002
 上記式(2)において、Kはクラス数、Mはデータ数、Tは時系列長である。また、kは行方向、lは列方向の添字(即ち、図6等に示すマトリクスにおける行番号及び列番号を示す添字)である。λは尤度比であり、上記数式(1)では、時刻tにおけるk行l列目の対数尤度比を表している。σは、シグモイド関数であり、尤度比λに作用する非線形関数として含まれている。
 式(2)は、正解クラスの行だけを考慮するため、すでに説明した式(1)と比べると、K行の合計を取り、1/2KでK行及び交代行列の平均をとる工程が含まれていない。また、式(1)におけるクロネッカーのδの部分が「1」となっている。
 (技術的効果)
 次に、第5実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図10で説明したように、第5実施形態に係る情報処理システム1では、学習部300が、シグモイド関数σを含む損失関数を利用する。このような損失関数を用いれば、上述した第1から第4実施形態において説明したように、適切に尤度比の算出に関する学習を行うことが可能となる。具体的には、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習が行える。
 第5実施形態で挙げた損失関数(即ち、式(1)及び式(2))は一例であり、シグモイド関数を利用して異なる損失関数を作成してもよい。また、シグモイド関数に代えて、他の非線形関数を用いて損失関数を作成してもよい。例えば、後述する実施形態のように、ロジスティック関数を含む損失関数を利用してもよい。
 <第6実施形態>
 第6実施形態に係る情報処理システム1について、図11を参照して説明する。なお、第6実施形態は、第5実施形態と同様に、上述した第1から第4実施形態で用いられる損失関数の具体例を説明するものであり、装置構成や動作の流れについては、第1から第4実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (ロジスティック関数を含む損失関数)
 まず、図11を参照しながら、第6実施形態に係る情報処理システム1で用いられる損失関数の概要について説明する。図11は、第6実施形態に係る情報処理システムで用いられるロジスティック関数の一例を示すグラフである。
 図11に示すように、第6実施形態に係る情報処理システム1では、学習部300がロジスティック関数を含む損失関数を用いて学習を行う。より具体的には、学習部300は、ロジスティック関数を、尤度比に作用する非線形関数として含む損失関数を用いて学習を行う。以下では、ロジスティック関数を含む損失関数の具体例を挙げて説明を進める。
 (すべての尤度比を考慮する例)
 1つ目の例として、第2実施形態のように、N×(N-1)パターンのすべての尤度比を考慮する場合の損失関数を挙げて説明する。なお、以下で示す式の尤度比は、対数尤度比(LLR:Log Likehood Ratio)であるものとする。
 N×(N-1)パターンのすべての尤度比を考慮する損失関数として、例えば下記式(3)が挙げられる。
Figure JPOXMLDOC01-appb-M000003
 上記式(3)において、Kはクラス数、Mはデータ数、Tは時系列長である。また、kは行方向、lは列方向の添字(即ち、図6等に示すマトリクスにおける行番号及び列番号を示す添字)である。δはクロネッカーのデルタであり、添字が一致する場合に「1」、それ以外の場合は「0」となる。λは尤度比であり、上記数式(3)では、時刻tにおけるk行l列目の対数尤度比を表している。logisticは、ロジスティック関数であり、尤度比λに作用する非線形関数として含まれている。
 式(1)における1/MTは、データ及び時系列方向の全てに対し平均をとるものである。1/2Kは、K行の平均をとる1/Kと、交代行列の平均をとる1/2とを乗じたものである。1/(K-1)は、K列から対角線の1列を差し引いた(K-1)列の平均をとるものである。
 式(3)の損失関数Lでは、k行が正解クラスであれば、クロネッカーのδが「1」となり、k行が不正解クラス(即ち、正解クラス以外のクラス)であれば、クロネッカーのδが「0」となる。この結果、k行が正解クラスであれば(言い換えれば、正解クラスが尤度比の分子にあれば)、クロネッカーのδを含む2つの項のうち、前の項が残り、後ろの項はゼロとなる。一方で、k行が不正解クラスであれば(言い換えれば、不正解クラスが尤度比の分子にあれば)、クロネッカーのδを含む2つの項のうち、前の項がゼロとなり、後ろの項が残る。
 (一部の尤度比を考慮する例)
 次に、2つ目の例として、第3及び第4実施形態のように、N×(N-1)パターンのうち一部の尤度比を考慮する場合の損失関数を挙げて説明する。以下では特に、第4実施形態のように、N×(N-1)パターンのうち、正解クラスが分子にある尤度比のみを考慮する例を挙げて説明する。
 N×(N-1)パターンのうち、正解クラスが分子にある尤度比のみを考慮する損失関数として、例えば下記式(4)が挙げられる。
Figure JPOXMLDOC01-appb-M000004
 上記式(4)において、Kはクラス数、Mはデータ数、Tは時系列長である。また、kは行方向、lは列方向の添字(即ち、図6等に示すマトリクスにおける行番号及び列番号を示す添字)である。λは尤度比であり、上記数式(1)では、時刻tにおけるk行l列目の対数尤度比を表している。logisticは、ロジスティック関数であり、尤度比λに作用する非線形関数として含まれている。
 式(4)は、正解クラスの行だけを考慮するため、すでに説明した式(3)と比べると、K行の合計を取り、1/2KでK行及び交代行列の平均をとる工程が含まれていない。また、式(3)におけるクロネッカーのδを含む2つの項のうち、前の項のみが残った形となっている。
 (技術的効果)
 次に、第6実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図11で説明したように、第6実施形態に係る情報処理システム1では、学習部300が、ロジスティックド関数を含む損失関数を利用する。このような損失関数を用いれば、上述した第1から第4実施形態において説明したように、適切に尤度比の算出に関する学習を行うことが可能となる。具体的には、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習が行える。また、第5実施形態で用いたシグモイド関数σは傾きがゼロに漸近するように変化する一方で(図10参照)、ロジスティック関数は傾きが一定に保たれたまま変化する(図11参照)。よって、ロジスティック関数を含む損失関数を用いた場合、損失関数の傾きを算出する処理(即ち、微分に相当する処理)において成分が消失せず、より適切な学習を行うことが可能である。
 第6実施形態で挙げた損失関数(即ち、式(3)及び式(4))は一例であり、ロジスティック関数を利用して異なる損失関数を作成してもよい。また、ロジスティック関数に代えて、他の非線形関数を用いて損失関数を作成してもよい。例えば、上述したシグモイド関数及びロジスティック関数のいずれとも異なる関数を含む損失関数を利用してもよい。
 <第7実施形態>
 第7実施形態に係る情報処理システム1について、図12及び図13を参照して説明する。なお、第3実施形態は、上述した第1から第6実施形態と一部の構成及び動作(具体的には、分類装置10の構成及び動作)が異なるのみで、その他の部分については、第1から第6実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (機能的構成)
 まず、図12を参照しながら、第7実施形態に係る情報処理システム1の機能的構成について説明する。図12は、第7実施形態に係る情報処理システムの機能的構成を示すブロック図である。なお、図12では、図2で示した各構成要素と同様の要素に同一の符号を付している。
 図12に示すように、第7実施形態に係る情報処理システム1では、分類装置10における尤度比算出部100が、第1算出部110と、第2算出部120とを備えている。なお、第1算出部110及び第2算出部120の各々は、例えば上述したプロセッサ11(図1参照)によって実現されてよい。
 第1算出部110は、系列データに含まれる2つの連続する要素に基づいて、個別尤度比を算出可能に構成されている。個別尤度比は、2つの連続する要素が属するクラスの尤もらしさを示す尤度比として算出される。第1算出部110は、例えばデータ取得部50から系列データに含まれる要素を逐次的に取得して、2つの連続する要素に基づく個別尤度比を順次算出してもよい。第1算出部110で算出された個別尤度比は、第2算出部120に出力される構成となっている。
 第2算出部120は、第1算出部110で算出される複数の個別尤度比に基づいて、統合尤度比を算出可能に構成されている。統合尤度比は、複数の個別尤度比の各々で考慮された複数の要素が属するクラスの尤もらしさを示す尤度比として算出される。言い換えれば、統合尤度比は、複数の要素を含んで構成される系列データが属するクラスの尤もらしさを示す尤度比として算出される。第2算出部120で算出された統合尤度比は、クラス分類部200に出力される構成となっている。クラス分類部200は、統合尤度比に基づいて系列データのクラス分類を行う。
 なお、第5実施形態に係る学習部300は、尤度比算出部100全体として(即ち、第1算出部110及び第2算出部120をまとめて)学習を行うようにしてもよいし、第1算出部110及び第2算出部120に対して別々に学習を行うようにしてもよい。或いは、学習部300は、第1算出部110のみ学習を行う第1の学習部、及び第2算出部120のみ学習を行う第2の学習部として別々に設けられてもよい。この場合、第1の学習部及び第2の学習部のいずれか一方のみが設けられてもよい。
 (分類動作の流れ)
 次に、図13を参照しながら、第7実施形態に係る情報処理システム1における分類装置10の動作(具体的には、学習後のクラス分類動作)の流れについて説明する。図13は、第7実施形態に係る情報処理システムにおける分類装置の動作の流れを示すフローチャートである。
 図13に示すように、分類装置10の動作が開始されると、まずデータ取得部50が、系列データに含まれる要素を取得する(ステップS21)。データ取得部50は、取得した系列データの要素を、第1算出部110に出力する。
 そして、第1算出部110は、取得された連続する2つの要素に基づいて、個別尤度比を算出する(ステップS22)。その後、第2算出部120が、第1算出部110で算出された複数の個別尤度比に基づいて、統合尤度比を算出する(ステップS23)。
 続いて、クラス分類部200が、算出された統合尤度比に基づいてクラス分類を行う(ステップS24)。クラス分類は、系列データが属する1つのクラスを決定するものであってもよいし、系列データが属する可能性の高い複数のクラスを決定するものであってもよい。クラス分類部200は、クラス分類の結果をディスプレイ等に出力するようにしてもよい。また、クラス分類部200は、クラス分類の結果を、スピーカ等を介して音声で出力するようにしてもよい。
 (技術的効果)
 次に、第7実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図12及び図13で説明したように、第7実施形態に係る情報処理システム1では、まず2つの要素に基づいて個別尤度比が算出された後、複数の個別尤度比に基づいて統合尤度比が算出される。このようにして算出される統合尤度比を用いれば、系列データが属するクラスを適切に選択することが可能となる。また、個別尤度比及び統合尤度比を算出する分類装置10においても、上述した各実施形態で説明した学習部300による学習を行うことで、クラス分類をより適切に行うことが可能となる。即ち、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習を行うことで、より適切な分類が行えるようになる。
 <第8実施形態>
 第8実施形態に係る情報処理システム1について、図14及び図15を参照して説明する。なお、第8実施形態は、上述した第7実施形態と一部の構成及び動作(具体的には、尤度比算出部100の構成及び動作)が異なるのみで、その他の部分については、第7実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (機能的構成)
 まず、図14を参照しながら、第8実施形態に係る情報処理システム1の機能的構成について説明する。図14は、第8実施形態に係る情報処理システムの機能的構成を示すブロック図である。なお、図14では、図2及び図12で示した各構成要素と同様の要素に同一の符号を付している。
 図14に示すように、第8実施形態に係る情報処理システム1では、分類装置10における尤度比算出部100が、第1算出部110と、第2算出部120とを備えている。第1算出部110は、個別尤度比算出部111と、第1記憶部とを備えて構成されている。第2算出部120は、統合尤度比算出部121と、第2記憶部122とを備えて構成されている。なお、個別尤度比算出部111及び統合尤度比算出部121の各々は、例えば上述したプロセッサ11(図1参照)によって実現されてよい。また、第1記憶部112及び第2記憶部122の各々は、例えば上述した記憶装置14(図1参照)によって実現されてよい。
 個別尤度比算出部111は、データ取得部50で逐次的に取得される要素のうち、連続する2つの要素に基づいて、個別尤度比を算出可能に構成されている。より具体的には、個別尤度比算出部111は、新たに取得した要素と、第1記憶部112に記憶された過去データに基づいて個別尤度比を算出する。第1記憶部112で記憶されている情報は、個別尤度比算出部111によって読み出し可能に構成されている。第1記憶部112が過去の個別尤度比を記憶している場合、個別尤度比算出部111は、記憶された過去の個別尤度比を読み出して、取得され要素を考慮した新たな個別尤度比を算出すればよい。一方、第1記憶部112が過去に取得された要素そのものを記憶している場合、個別尤度比算出部111は、記憶された過去の要素から過去の個別尤度比を算出して、新たに取得された要素に対する尤度比を算出すればよい。
 統合尤度比算出部121は、複数の個別尤度比に基づいて、統合尤度比を算出可能に構成されている。統合尤度比算出部121は、個別尤度比算出部111で算出された個別尤度比と、第2記憶部122に記憶された過去の統合尤度比とを用いて、新たな統合尤度比を算出する。第2記憶部122で記憶されている情報(即ち、過去の統合尤度比)は、統合尤度比算出部121によって読み出し可能に構成されている。
 <尤度比算出動作の流れ>
 次に、図15を参照しながら、第8実施形態に係る情報処理システム1における尤度比算出動作(即ち、尤度比算出部100の動作)の流れについて説明する。図15は、第8実施形態に係る情報処理システムにおける尤度比算出部の動作の流れを示すフローチャートである。
 図15に示すように、尤度比算出部100による尤度比算出動作が開始されると、まず第1算出部110における個別尤度比算出部111が、第1記憶部112から過去データを読み出す(ステップS31)。過去データは、例えばデータ取得部50で今回取得された要素の1つ前に取得された要素の個別尤度比算出部111での処理結果(言い換えれば、1つ前の要素に対して算出された個別尤度比)であってよい。或いは、過去データは、取得で取得された要素の1つ前に取得された要素そのものであってもよい。
 続いて、個別尤度比算出部111は、データ取得部50で取得された要素と、第1記憶部112から読みだした過去データに基づいて、新たな個別尤度比(即ち、データ取得部50で今回取得された要素に対する個別尤度比)を算出する(ステップS32)。個別尤度比算出部111は、算出した個別尤度比を、第2算出部120に出力する。個別尤度比算出部111は、算出した個別尤度比を、第1記憶部112に記憶してもよい。
 続いて、第2算出部120における統合尤度比算出部121が、第2記憶部122から過去の統合尤度比を読み出す(ステップS33)。過去の統合尤度比は、例えばデータ取得部50で今回取得された要素の1つ前に取得された要素についての、統合尤度比算出部121での処理結果(言い換えれば、1つ前の要素に対して算出された統合尤度比)であってよい。
 続いて、統合尤度比算出部121は、個別尤度比算出部111で算出された尤度比と、第2記憶部122から読みだした過去の統合尤度比に基づいて、新たな統合尤度比(即ち、データ取得部50で今回取得された要素に対する統合尤度比)を算出する(ステップS34)。統合尤度比算出部121は、算出した統合尤度比を、クラス分類部200に出力する。統合尤度比算出部121は、算出した統合尤度比を、第2記憶部122に記憶してもよい。
 (技術的効果)
 次に、第8実施形態に係る情報処理システム1によって得られる技術的効果について説明する。
 図14及び図15で説明したように、第8実施形態に係る情報処理システム1では、過去の個別尤度比を用いて個別尤度比が算出された後、過去の統合尤度比を用いて統合尤度比が算出される。このようにして算出される統合尤度比を用いれば、系列データが属するクラスを適切に選択することが可能となる。また、過去のデータを用いて個別尤度比及び統合尤度比を算出する分類装置10においても、上述した各実施形態で説明した学習部300による学習を行うことで、クラス分類をより適切に行うことが可能となる。即ち、クラスが不正解である場合のペナルティが大きくなり、クラスが正解である場合のペナルティが小さくなるような学習を行うことで、より適切な分類が行えるようになる。
 <第9実施形態>
 第9実施形態に係る情報処理システム1について、図16を参照して説明する。なお、第9実施形態は、上述した第1から第7実施形態と一部の動作(具体的には、クラス分類部200の動作)が異なるのみで、その他の部分については、第1から第7実施形態と同様であってよい。このため、以下では、すでに説明した各実施形態と異なる部分について詳しく説明し、他の重複する部分については適宜説明を省略するものとする。
 (分類動作の流れ)
 まず、図16を参照しながら、第9実施形態に係る情報処理システム1における分類装置10の動作(具体的には、学習後のクラス分類動作)の流れについて説明する。図16は、第9実施形態に係る情報処理システムにおける分類装置の動作の流れを示すフローチャートである。なお、図16では、図3で説明した処理と同様の処理に同一の符号を付している。
 図16に示すように、分類装置10の動作が開始されると、まずデータ取得部50が、系列データに含まれる要素を取得する(ステップS11)。データ取得部50は、取得した系列データの要素を、尤度比算出部100に出力する。そして、尤度比算出部100は、取得された2つ以上の要素に基づいて尤度比を算出する(ステップS12)。
 続いて、クラス分類部200が、算出された尤度比に基づいてクラス分類を行うが、第9実施形態では特に、クラス分類部200が、系列データが属する可能性がある複数のクラスを選択して出力する(ステップS41)。即ち、クラス分類部200は、系列データが属する1つのクラスを決定するのではなく、系列データが属する可能性が高い複数のクラスを決定する。より具体的には、クラス分類部200は、分類候補として用意されたn個のクラス(ただし、nは自然数)の中から、k個のクラス(ただし、kはn以下の自然数)を選択する処理を実行する。
 クラス分類部200は、系列データが属する可能性があるk個のクラスに関する情報を、ディスプレイ等に出力するようにしてもよい。また、クラス分類部200は、系列データが属する可能性があるk個のクラスに関する情報を、スピーカ等を介して音声で出力するようにしてもよい。
 クラス分類部200は、系列データが属する可能性があるk個のクラスに関する情報を出力する際に、それらを並び替えて出力してもよい。例えば、クラス分類部200は、k個のクラスに関する情報を尤度比の高い順に並び替えて出力してもよい。或いは、クラス分類部200は、k個のクラスに関する情報の各々をクラス毎に異なる態様で出力するようにしてもよい。例えば、クラス分類部200は、尤度比の高いクラスを強調した表示態様で出力する一方で、尤度比の低いクラスを強調しない表示態様で出力するようにしてもよい。強調表示する場合、例えば表示する大きさや色を変えたりしてもよいし、表示するものに動きを与えてもよい。
 (具体的な活用例)
 上述したn個のクラスの中からk個のクラスを出力する構成について、具体的な適用例をいくつか挙げて説明する。
 (商品の提案)
 第9実施形態に係る情報処理システム1は、ウェブ上のショッピングサイトにおいて、ユーザが興味を持ちそうな商品の提案に用いられてもよい。具体的には、情報処理システム1は、取扱商品であるn個の商品(即ち、n個のクラス)の中から、ユーザが興味を持ちそうなk個の商品(即ち、k個のクラス)を選択して、ユーザに対して出力するようにしてもよい(なお、kはnより小さい数である)。この場合、入力される系列データの一例として、過去の購入履歴や閲覧履歴等が挙げられる。
 同様に、デジタルサイネージ等における商品や店舗の提案に用いられてもよい。デジタルサイネージでは、搭載されたカメラによってユーザの画像を撮像可能となっていることがある。この場合、ユーザの画像からユーザの感情を推定して、その感情に応じた店舗や商品を提案するようにしてもよい。また、ユーザの画像からユーザの視線を推定して(即ち、ユーザが見ている部分を推定して)、ユーザが興味を持ちそうな店舗や商品を提案するようにしてもよい。或いは、ユーザの画像からユーザの属性(例えば、性別や年齢等)を推定して、ユーザが興味を持ちそうな店舗や商品を提案するようにしてもよい。なお、上記のようにユーザに関する情報を推定する場合、推定した情報に応じてn個のクラスに重み付けを行ってよい。
 (犯罪捜査)
 第9実施形態に係る情報処理システム1は、犯罪捜査に用いることもできる。例えば、複数の容疑者の中から真犯人を見つける場合、その中から最も犯人である可能性の高い1人だけを選択すると、その選択が間違っていた場合に大きな問題が生ずる。しかるに、本実施形態に係る情報処理システム1では、犯人である可能性が高い上位K人の容疑者を選択して出力することができる。具体的には、複数の容疑者の各々に関する情報を要素として含む系列データから、犯人である可能性が高い上位k人に対応するクラスを選択して出力するようにすればよい。このようにすれば、例えば犯人である可能性が高い複数の容疑者を捜査対象として、適切に真犯人を見つけ出すことが可能となる。
 (レーダ画像分析)
 第9実施形態に係る情報処理システム1は、レーダ画像の分析に適用することもできる。レーダ画像は、その性質上、鮮明度が低いものが多いため、例えばその画像に写っているものが何であるのか、機械のみで正確に判定することが難しい。しかるに本実施形態に係る情報処理システム1では、レーダ画像に写っている可能性が高いk個の候補を選択して出力することができる。よって、まずはk個の候補を出力し、その中からユーザ自身で判断してもらうことが可能である。例えば、港のレーダ画像に写っているものとして、「犬」、「猫」、「船」、及び「戦車」が候補として挙げられれば、ユーザは港に関連の高い「船」がレーダ画像に写っていると容易に判断できる。
 なお、上述した適用例は一例であり、n個の候補の中から、k個のものを選択することが要求されるような状況であれば、本実施形態に係る情報処理システム1を適用することで、有益な効果を奏することが可能である。
 上述した各実施形態の機能を実現するように該実施形態の構成を動作させるプログラムを記録媒体に記録させ、該記録媒体に記録されたプログラムをコードとして読み出し、コンピュータにおいて実行する処理方法も各実施形態の範疇に含まれる。すなわち、コンピュータ読取可能な記録媒体も各実施形態の範囲に含まれる。また、上述のプログラムが記録された記録媒体はもちろん、そのプログラム自体も各実施形態に含まれる。
 記録媒体としては例えばフロッピー(登録商標)ディスク、ハードディスク、光ディスク、光磁気ディスク、CD-ROM、磁気テープ、不揮発性メモリカード、ROMを用いることができる。また該記録媒体に記録されたプログラム単体で処理を実行しているものに限らず、他のソフトウェア、拡張ボードの機能と共同して、OS上で動作して処理を実行するものも各実施形態の範疇に含まれる。
 この開示は、請求の範囲及び明細書全体から読み取ることのできる発明の要旨又は思想に反しない範囲で適宜変更可能であり、そのような変更を伴う情報処理システム、情報処理方法、及びコンピュータプログラムもまたこの開示の技術思想に含まれる。
 <付記>
 以上説明した実施形態に関して、更に以下の付記のようにも記載されうるが、以下には限られない。
 (付記1)
 付記1に記載の情報処理システムは、系列データに含まれる複数の要素を取得する取得手段と、前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出する算出手段と、前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類する分類手段と、前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行う学習手段とを備えることを特徴とする情報処理システムである。
 (付記2)
 付記2に記載の情報処理システムは、前記学習手段は、前記系列データの分類候補であるN個(ただし、Nは自然数)のクラスのうち、一のクラスに属する尤もらしさを示す尤度を分母とし、他のクラスに属する尤もらしさを示す尤度を分子とするN×(N-1)パターンの前記尤度比を考慮した損失関数を用いて、前記学習を行うことを特徴とする付記1に記載の情報処理システムである。
 (付記3)
 付記3に記載の情報処理システムは、前記学習手段は、前記N×(N-1)パターンのうち一部の前記尤度比を考慮した損失関数を用いて、前記学習を行うことを特徴とする付記2に記載の情報処理システムである。
 (付記4)
 付記4に記載の情報処理システムは、前記学習手段は、前記N×(N-1)パターンのうち、前記正解クラスが分子にある尤度比を考慮した損失関数を用いて、前記学習を行うことを特徴とする付記3に記載の情報処理システムである。
 (付記5)
 付記5に記載の情報処理システムは、前記損失関数は、前記尤度比に作用する非線形関数としてシグモイド関数を含んでいることを特徴とする付記1から4のいずれか一項に記載の情報処理システムである。
 (付記6)
 付記6に記載の情報処理システムは、前記損失関数は、前記尤度比に作用する非線形関数としてロジスティック関数を含んでいることを特徴とする付記1から4のいずれか一項に記載の情報処理システムである。
 (付記7)
 付記7に記載の情報処理システムは、前記尤度比は、前記系列データに含まれる2つの連続する要素に基づいて算出される個別尤度比を複数考慮して算出される統合尤度比であることを特徴とする付記1から6のいずれか一項に記載の情報処理システムである。
 (付記8)
 付記8に記載の情報処理システムは、前記取得手段は、系列データに含まれる複数の要素を逐次的に取得し、前記算出手段は、新たに取得した要素に基づいて算出した前記個別尤度比と、過去に算出した前記統合尤度比とを用いて、新たな前記統合尤度比を算出することを特徴とする付記7に記載の情報処理システムである。
 (付記9)
 付記9に記載の情報処理方法は、系列データに含まれる複数の要素を取得し、前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出し、前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類し、前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行うことを特徴とする情報処理方法である。
 (付記10)
 付記10に記載のコンピュータプログラムは、系列データに含まれる複数の要素を取得し、前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出し、前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類し、前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行うようにコンピュータを動作させることを特徴とするコンピュータプログラムである。
 (付記11)
 付記11に記載の記録媒体は、付記10に記載のコンピュータプログラムが記録されていることを特徴とする記録媒体である。
 1 情報処理システム
 11 プロセッサ
 14 記憶装置
 10 分類装置
 50 データ取得部
 100 尤度比算出部
 110 第1算出部
 111 個別尤度比算出部
 112 第1記憶部
 120 第2算出部
 121 統合尤度比算出部
 122 第2記憶部
 200 クラス分類部
 300 学習部

Claims (10)

  1.  系列データに含まれる複数の要素を取得する取得手段と、
     前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出する算出手段と、
     前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類する分類手段と、
     前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行う学習手段と
     を備えることを特徴とする情報処理システム。
  2.  前記学習手段は、前記系列データの分類候補であるN個(ただし、Nは自然数)のクラスのうち、一のクラスに属する尤もらしさを示す尤度を分母とし、他のクラスに属する尤もらしさを示す尤度を分子とするN×(N-1)パターンの前記尤度比を考慮した損失関数を用いて、前記学習を行うことを特徴とする請求項1に記載の情報処理システム。
  3.  前記学習手段は、前記N×(N-1)パターンのうち一部の前記尤度比を考慮した損失関数を用いて、前記学習を行うことを特徴とする請求項2に記載の情報処理システム。
  4.  前記学習手段は、前記N×(N-1)パターンのうち、前記正解クラスが分子にある尤度比を考慮した損失関数を用いて、前記学習を行うことを特徴とする請求項3に記載の情報処理システム。
  5.  前記損失関数は、前記尤度比に作用する非線形関数としてシグモイド関数を含んでいることを特徴とする請求項1から4のいずれか一項に記載の情報処理システム。
  6.  前記損失関数は、前記尤度比に作用する非線形関数としてロジスティック関数を含んでいることを特徴とする請求項1から4のいずれか一項に記載の情報処理システム。
  7.  前記尤度比は、前記系列データに含まれる2つの連続する要素に基づいて算出される個別尤度比を複数考慮して算出される統合尤度比であることを特徴とする請求項1から6のいずれか一項に記載の情報処理システム。
  8.  前記取得手段は、系列データに含まれる複数の要素を逐次的に取得し、
     前記算出手段は、新たに取得した要素に基づいて算出した前記個別尤度比と、過去に算出した前記統合尤度比とを用いて、新たな前記統合尤度比を算出する
     ことを特徴とする請求項7に記載の情報処理システム。
  9.  系列データに含まれる複数の要素を取得し、
     前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出し、
     前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類し、
     前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行う
     ことを特徴とする情報処理方法。
  10.  系列データに含まれる複数の要素を取得し、
     前記複数の要素のうち少なくとも2つの連続する要素に基づいて、前記系列データが属するクラスの尤もらしさを示す尤度比を算出し、
     前記尤度比に基づいて、分類候補である複数のクラスのうち少なくとも1つのクラスに前記系列データを分類し、
     前記系列データが属する正解クラスが前記尤度比の分子にある場合に前記尤度比が大きくなり、前記正解クラスが前記尤度比の分母にある場合に前記尤度比が小さくなる損失関数を用いて、前記尤度比の算出に関する学習を行う
     ようにコンピュータを動作させることを特徴とするコンピュータプログラム。
PCT/JP2020/048472 2020-12-24 2020-12-24 情報処理システム、情報処理方法、及びコンピュータプログラム WO2022137439A1 (ja)

Priority Applications (3)

Application Number Priority Date Filing Date Title
PCT/JP2020/048472 WO2022137439A1 (ja) 2020-12-24 2020-12-24 情報処理システム、情報処理方法、及びコンピュータプログラム
US18/267,730 US20240054400A1 (en) 2020-12-24 2020-12-24 Information processing system, information processing method, and computer program
JP2022570890A JPWO2022137439A1 (ja) 2020-12-24 2020-12-24

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2020/048472 WO2022137439A1 (ja) 2020-12-24 2020-12-24 情報処理システム、情報処理方法、及びコンピュータプログラム

Publications (1)

Publication Number Publication Date
WO2022137439A1 true WO2022137439A1 (ja) 2022-06-30

Family

ID=82159249

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2020/048472 WO2022137439A1 (ja) 2020-12-24 2020-12-24 情報処理システム、情報処理方法、及びコンピュータプログラム

Country Status (3)

Country Link
US (1) US20240054400A1 (ja)
JP (1) JPWO2022137439A1 (ja)
WO (1) WO2022137439A1 (ja)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH1138993A (ja) * 1997-07-16 1999-02-12 Oki Electric Ind Co Ltd Hmmパラメータ学習方法と装置
JP2007114413A (ja) * 2005-10-19 2007-05-10 Toshiba Corp 音声非音声判別装置、音声区間検出装置、音声非音声判別方法、音声区間検出方法、音声非音声判別プログラムおよび音声区間検出プログラム
JP2010003117A (ja) * 2008-06-20 2010-01-07 Toyota Central R&D Labs Inc 対象物判定装置及びプログラム
JP2011027910A (ja) * 2009-07-23 2011-02-10 Kddi Corp パターン認識方法および装置ならびにパターン認識プログラムおよびその記録媒体
JP2017151679A (ja) * 2016-02-24 2017-08-31 コニカミノルタ株式会社 識別装置及び識別プログラム

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH1138993A (ja) * 1997-07-16 1999-02-12 Oki Electric Ind Co Ltd Hmmパラメータ学習方法と装置
JP2007114413A (ja) * 2005-10-19 2007-05-10 Toshiba Corp 音声非音声判別装置、音声区間検出装置、音声非音声判別方法、音声区間検出方法、音声非音声判別プログラムおよび音声区間検出プログラム
JP2010003117A (ja) * 2008-06-20 2010-01-07 Toyota Central R&D Labs Inc 対象物判定装置及びプログラム
JP2011027910A (ja) * 2009-07-23 2011-02-10 Kddi Corp パターン認識方法および装置ならびにパターン認識プログラムおよびその記録媒体
JP2017151679A (ja) * 2016-02-24 2017-08-31 コニカミノルタ株式会社 識別装置及び識別プログラム

Also Published As

Publication number Publication date
US20240054400A1 (en) 2024-02-15
JPWO2022137439A1 (ja) 2022-06-30

Similar Documents

Publication Publication Date Title
CN109697434B (zh) 一种行为识别方法、装置和存储介质
US10832069B2 (en) Living body detection method, electronic device and computer readable medium
US20230041233A1 (en) Image recognition method and apparatus, computing device, and computer-readable storage medium
US11093789B2 (en) Method and apparatus for object re-identification
CN106203376B (zh) 人脸关键点定位方法及装置
CN107690657B (zh) 根据影像发现商户
CN102207950B (zh) 电子装置和图像处理方法
CN112016538B (zh) 视频处理方法、装置、计算机设备和存储介质
CN110175609B (zh) 界面元素检测方法、装置及设备
EP3623998A1 (en) Character recognition
CN111767883A (zh) 一种题目批改方法及装置
CN110210480B (zh) 文字识别方法、装置、电子设备和计算机可读存储介质
US10937428B2 (en) Pose-invariant visual speech recognition using a single view input
CN114120349B (zh) 基于深度学习的试卷识别方法及系统
CN112488923A (zh) 图像超分辨率重建方法、装置、存储介质及电子设备
CN110858316A (zh) 对时间序列图像数据进行分类
WO2022157973A1 (ja) 情報処理システム、情報処理方法、及びコンピュータプログラム
CN114005019B (zh) 一种翻拍图像识别方法及其相关设备
CN112115862A (zh) 一种结合密度估计的拥挤场景行人检测方法
CN111405314A (zh) 一种信息处理方法、装置、设备及存储介质
WO2022137439A1 (ja) 情報処理システム、情報処理方法、及びコンピュータプログラム
CN112464810A (zh) 一种基于注意力图的吸烟行为检测方法及装置
CN112529078A (zh) 一种业务处理方法、装置及设备
JP7293658B2 (ja) 情報処理装置、情報処理方法及びプログラム
Dey et al. Mood recognition in online sessions using machine learning in realtime

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20966928

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2022570890

Country of ref document: JP

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 18267730

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20966928

Country of ref document: EP

Kind code of ref document: A1