Similarity calculation device, similarity calculation method, and computer-readable recording medium

Abstract

A similarity calculation device (1) has a similarity calculation section (19) that performs transfer learning using a transfer source data candidate (2) and transfer destination data (3). The similarity calculation section (19) calculates a similarity of a transfer source data distribution and a transfer destination data distribution from a transfer source extreme value group including a transfer source extreme value indicating an extreme value possessed by the transfer source data distribution and a transfer destination extreme value group including a transfer destination extreme value indicating an extreme value possessed by the transfer destination data distribution, the transfer source data distribution indicating a distribution of a feature quantity of the transfer source data candidate (2), the transfer destination data distribution indicating a distribution of a feature quantity of the transfer destination data (3).

Description

Technical Field

[0001] This invention relates to a similarity calculation device, a similarity calculation method, and a computer-readable recording medium. Background Technology

[0002] Solutions utilizing AI (Artificial Intelligence) on IoT (Internet of Things) devices are increasing. Specific examples include (1) control of IoT home appliances such as air conditioners and lighting, (2) fault diagnosis of manufacturing equipment, (3) image-based inspection of products on production lines, (4) dynamic image-based suspicious person intrusion detection at building entrances, and (5) EMS (Energy Management System).

[0003] When utilizing AI on a per-IoT device basis, it is often difficult to ensure a sufficient amount of training data for the learning process. Therefore, it is necessary to perform learning processes efficiently using less training data. One method for learning with less training data is called transfer learning. In transfer learning, training data and the learning model are transferred from an environment where the training data was collected.

[0004] In transfer learning, to determine the transfer source, each candidate data point is evaluated to determine its suitability as a transfer source. If a candidate is confirmed through evaluation as a valid "positive transfer," it is designated as the transfer source. When the training data consists of images, the sufficient similarity between the candidate image and the destination image is often a sufficient condition for "positive transfer."

[0005] As a method for calculating image similarity, Patent Document 1 discloses a method for calculating the similarity between a target image and a source image using a certain evaluation function, such as a method for calculating the similarity by calculating the color histogram of each image, a method for calculating the similarity by using the gradient direction of brightness as a feature, and a method for setting the number of similar features of local features detected from the input image and the learning image as the evaluation value.

[0006] Existing technical documents

[0007] Patent documents

[0008] Patent Document 1: Japanese Patent Application Publication No. 2019-087229 Summary of the Invention

[0009] The problem that the invention aims to solve

[0010] According to the method disclosed in Patent Document 1, there is a problem of requiring a large amount of computation.

[0011] The purpose of this invention is to determine the transfer source with less computation in transfer learning.

[0012] Methods for solving problems

[0013] The similarity calculation apparatus of the present invention performs transfer learning using migration source data candidates and migration destination data. The apparatus includes a similarity calculation unit that calculates the similarity between the migration source data distribution and the migration destination data distribution based on a migration source extreme value set including migration source extreme values ​​representing extreme values ​​in the migration source data distribution and a migration destination extreme value set including migration destination extreme values ​​representing extreme values ​​in the migration destination data distribution. The migration source data distribution represents the distribution of feature quantities of the migration source data candidates, and the migration destination data distribution represents the distribution of feature quantities of the migration destination data.

[0014] Invention Effects

[0015] According to the present invention, the transfer source can be determined with less computation in transfer learning. Attached Figure Description

[0016] Figure 1 This is a structural example of the similarity calculation device 1 in Implementation 1.

[0017] Figure 2 This is a specific example of the feature data distribution 102S in Implementation Method 1.

[0018] Figure 3 This is a specific example of dataset 105 in implementation method 1.

[0019] Figure 4 This is an example of the hardware structure of the similarity calculation device 1 in Implementation 1.

[0020] Figure 5 This is a diagram illustrating the outline of the operation of the similarity calculation device 1 in Embodiment 1.

[0021] Figure 6 This is a flowchart illustrating the actions of the similarity calculation device 1 in Embodiment 1 related to the learning phase.

[0022] Figure 7 This is a diagram illustrating the actions of the similarity calculation device 1 in the learning phase of embodiment 1. (a) is an image showing candidates for migration source data, (b) is an image of (a) in grayscale, and (c) is the feature data distribution corresponding to (b).

[0023] Figure 8This is a flowchart illustrating the actions of the similarity calculation device 1 in Embodiment 1 related to the reasoning stage.

[0024] Figure 9 This is a diagram illustrating the operation of the similarity calculation device 1 in Embodiment 1.

[0025] Figure 10 This is a hardware structure example of the similarity calculation device 1, which is a variation of implementation 1.

[0026] Figure 11 These are diagrams illustrating the general operation of the similarity calculation device 1 in Embodiment 2. (a) is a diagram illustrating the operation before the coordinate system is transformed, and (b) is a diagram illustrating the operation after the coordinate system is transformed.

[0027] Figure 12 This is a structural example of the similarity calculation device 1 in Embodiment 2.

[0028] Figure 13 This is a flowchart illustrating the actions of the similarity calculation device 1 in Embodiment 2 related to the learning phase.

[0029] Figure 14 This is a flowchart illustrating the actions of the similarity calculation device 1 in Embodiment 2 related to the reasoning stage.

[0030] Figure 15 These are diagrams illustrating the general operation of the similarity calculation device 1 in Embodiment 3. (a) is a diagram illustrating the case without a tolerance range, and (b) is a diagram illustrating the case with a tolerance range.

[0031] Figure 16 This is a structural example of the similarity calculation device 1 in embodiment 3.

[0032] Figure 17 This is a diagram illustrating the operation of the similarity calculation device 1 in Embodiment 3.

[0033] Figure 18 These are diagrams illustrating the general operation of the similarity calculation device 1 in Embodiment 4. (a) is a diagram illustrating the case without a tolerance range, and (b) is a diagram illustrating the case with a tolerance range.

[0034] Figure 19 This is a structural example of the similarity calculation device 1 in embodiment 4. Detailed Implementation

[0035] In the description and accompanying drawings of the embodiments, the same reference numerals are used to label the same elements and corresponding elements. Descriptions of elements labeled with the same reference numerals are appropriately omitted or simplified. Arrows in the figures primarily indicate data flow or processing flow. Furthermore, the term "section" may be appropriately replaced with "circuit," "process," "processing," or "line."

[0036] Implementation Method 1

[0037] Hereinafter, this embodiment will be described in detail with reference to the accompanying drawings.

[0038] ***Structure Description***

[0039] Figure 1 The diagram shows a structural example of the similarity calculation device 1 according to this embodiment. As shown in the figure, the similarity calculation device 1 includes a feature extraction unit 11, a data distribution calculation unit 12, an extreme value calculation unit 13, a grouping unit 14, a dataset storage unit 15, a model generation unit 16, a classifier storage unit 17, a model determination unit 18, a similarity calculation unit 19, and a data determination unit 20.

[0040] Furthermore, the similarity calculation device 1 uses the candidate source data 2 and the destination data 3 for transfer learning. Additionally, the T at the end of the label and the t superscript indicate correlation with the destination data distribution, while the S at the end of the label and the s superscript indicate correlation with the source data distribution.

[0041] The feature extraction unit 11 uses the candidate source data 2 to generate a feature vector 101S representing the feature values ​​of the candidate source data 2, and uses the destination data 3 to generate a feature vector 101T representing the feature values ​​of the destination data 3. The candidate source data 2 and the destination data 3 can be data received from other devices or data stored in the similarity calculation device 1. Furthermore, the candidate source data 2 is a candidate for the source data, and at least one exists. It is assumed that there is one destination data 3. The data form of the feature values ​​obtained by the feature extraction unit 11 may not be in vector form. Additionally, the feature vector 101 is referred to collectively as feature vector 101S and feature vector 101T. As a specific example, feature vector 101 is an arrangement of pixel values.

[0042] As a specific example, the feature extraction unit 11 calculates at least one of the following as feature quantities: data obtained by summarizing the source data candidate 2 or destination data 3 through principal component analysis; statistics such as the mean, variance, or correlation coefficient related to the source data candidate 2 or destination data 3; values ​​obtained by transforming the source data candidate 2 or destination data 3 using a mapping of the same kind, such as a histogram or Fourier transform; and gradient vectors used as coefficients or their changes, obtained by approximating the source data candidate 2 or destination data 3 as a function using regression analysis, etc. The term "histogram" sometimes refers to data that can be represented by a histogram. Furthermore, it is assumed that the feature quantities calculated by the feature extraction unit 11 sufficiently reflect the characteristics of the source data candidate 2 or destination data 3.

[0043] The data distribution calculation unit 12 generates a feature data distribution 102S using feature vector 101S and a feature data distribution 102T using feature vector 101T. Feature data distribution 102S represents the distribution of feature quantities of the migration source data candidate 2, and is also called the migration source data distribution; at least one such distribution exists. Each migration source data candidate 2 corresponds one-to-one with each feature data distribution 102S. Feature data distribution 102T represents the distribution of feature quantities of the migration destination data 3, and is also called the migration destination data distribution. Furthermore, feature data distribution 102 is referred to collectively as feature data distribution 102S and feature data distribution 102T. Feature data distribution 102 is typically a function, but can also be a discrete distribution or a continuous distribution.

[0044] The extreme value calculation unit 13 uses the characteristic data distribution 102S to calculate the extreme value data 103S. The extreme value data 103S is determined by the data ID (Identification). s Extreme value coordinate system G s And the number of extreme values ​​E s The composition, also known as the calculation result. Data ID s This is the identifier for candidate migration source data 2, also known as the migration source data ID. Extreme value coordinate group G s It is a set of coordinates of the extreme values ​​of the feature data distribution 102S, also known as the migration source extreme value coordinate set or migration source extreme value set, and at least one such set exists. Each feature data distribution 102S and each extreme value coordinate set G... s A one-to-one correspondence. Extreme value coordinate group G s The included extreme values ​​are also called migration source extreme values. Each migration source extreme value corresponds to an identifier identifying migration source data candidate 2, which corresponds to a feature data distribution 102S having extreme values ​​corresponding to the migration source extreme values. The number of extreme values ​​E s It is the extreme value coordinate system G s The number of extreme values ​​included.

[0045] Furthermore, the extreme value calculation unit 13 uses the characteristic data distribution 102T to calculate the extreme value data 103T. The extreme value data 103T is derived from the extreme value coordinate group G. t And the number of extreme values ​​E t The composition, also known as the calculation result, is the extreme value coordinate system G. t It is a set of coordinates of the extreme values ​​of the feature data distribution 102T, also known as the destination extreme value coordinate set or destination extreme value set. The number of extreme values ​​E t It is the extreme value coordinate system G t The number of extreme values ​​included. The extreme value data 103T can also contain identifiers identifying the migration destination data 3. Set the number of extreme values ​​to the number of extreme values ​​E. sAnd the number of extreme values ​​E t A general term.

[0046] The extreme value calculation unit 13 receives the migration source data distribution and the migration destination data distribution, uses the migration source data distribution to calculate the extreme value of the migration source data distribution, and uses the migration destination data distribution to calculate the extreme value of the migration destination data distribution.

[0047] In the characteristic data distribution 102S, y = f s (x) and the number of extreme values ​​E s When the value is n, the extreme value coordinate system G s Included coordinates (x) i s y i s As defined in [Equation 1]. [Equation 1] is also the extreme value coordinate system G. s Specific examples. Additionally, the extreme value coordinate system G... s The dimension of each coordinate contained herein is not limited to 2; however, for ease of explanation, we assume that the dimension of each coordinate is 2. Extreme value coordinate set G t Defined in the same way as in [Equation 1]. Furthermore, let the variable s be an element of the set S consisting of all assigned feature data distributions 102S. That is, s∈S.

[0048] [Formula 1]

[0049]

[0050] Figure 2 A specific example of the characteristic data distribution 102S is shown. This figure illustrates the extreme value coordinate set G for the case where n is 6, as shown in [Equation 1]. s Specific examples of each extreme value included.

[0051] The grouping unit 14 uses the extreme value data 103S to generate learning data 104, and saves the generated learning data 104 to the dataset storage unit 15. Learning data 104 contains training data and supervised data. In learning data 104, the training data is the extreme value coordinate set G. s The included extreme value coordinates, the supervision data are the data IDs corresponding to the feature data distribution 102S which has extreme value coordinates as training data. s Supervisory data is also labeled. Labels are also identifiers for identifying candidate 2 of the migration source data.

[0052] The grouping unit 14 utilizes the number E of extreme values ​​corresponding to the learning data 104. s After classifying the learning data 104, the learning data 104 is saved to the dataset storage unit 15. The number E of extreme values ​​corresponding to the learning data 104... sIt is the set of extreme coordinates G, which includes the extreme coordinates contained in the learning data 104. s The number of extreme values ​​E contained s .

[0053] Dataset storage section 15 stores datasets 105-1 to 105-N; it is also referred to as dataset 1 to N storage section. N represents the number of extreme values, E. s The constant of the maximum value can also be changed dynamically. The dataset 105-n (1≦n≦N, where n is an integer) is the number of extreme values ​​E respectively. s Let n be the set of learning data (104 elements) corresponding to the feature data distribution 102S of n. That is, n is the number of extreme values ​​E. s The corresponding variables. Additionally, at least some of the datasets 105-1 to 105-N can be empty data.

[0054] Each migration source extreme value is associated with an identifier for a migration source data candidate 2, which corresponds to a feature data distribution 102S having extreme values ​​corresponding to the migration source extreme values. Each migration source extreme value belongs to at least one of the datasets 105 based on the number of migration source extreme values ​​contained in each migration source extreme value group. The dataset 105 is prepared for each value represented by the number of extreme values. In this embodiment, the dataset 105 contains migration source extreme values ​​corresponding to the number of migration source extreme values, each corresponding to a dataset 105. The number of migration source extreme values ​​corresponding to dataset 105-n is n. When the number of migration source extreme values ​​is n, the migration source extreme values ​​corresponding to the number of migration source extreme values ​​are the migration source extreme values ​​possessed by the feature data distribution 102S having n migration source extreme values.

[0055] Equation 2 shows the number of extreme values ​​E s For n(=|{(x i s y i s The distribution of all feature data of )}|) is a specific example of the dataset 105-n generated by 102S.

[0056] [Equation 2]

[0057]

[0058] Here, the dataset 105-n is represented as a set g. n The extreme value coordinate group G s The included coordinates are represented as (x i s y i s ), will data ID sThe manifestation is c s .

[0059] Figure 3 This shows a specific example of dataset 105. The extreme value coordinate group G is shown at the top of the graph. s G s1 G s2 ... G sn It is the coordinate system of each extreme value G s Here, the superscript subscript indicates that the coordinates of each extreme value set G are related to the superscript subscript. s The data ID corresponding to the feature data distribution of the included extreme value coordinates 102S s The label. Furthermore, the lower part of the figure shows the extreme value coordinate group G, which is the same as the upper part. s The corresponding dataset is 105. (This is related to G...) s1 and G sn The number of corresponding extreme values ​​E s The value is 3, therefore, G s1 and G sn The elements are contained within the elements contained in dataset 105-3, and are related to G. s2 The number of corresponding extreme values ​​E s The value is 1, therefore, G s2 The elements are contained in the elements contained in dataset 105-1.

[0060] The model generation unit 16 learns from the dataset 105 and generates a classifier 106 as a learning model. Specifically, the model generation unit 16 obtains the dataset 105-n from the dataset storage unit 15 and uses the obtained dataset 105-n to generate a classifier 106-n. That is, the model generation unit 16 generates the classifier 106-n according to the number of extreme values ​​E. s Classifier 106 is generated. Alternatively, if dataset 105-n is empty, model generation unit 16 may not generate classifier 106-n. Classifier 106, as the learning model, is a model corresponding to dataset 105, and is used to estimate the identifiers corresponding to the assigned coordinate sets; it is a model used to classify the extreme values ​​of the feature data distribution 102T. Specifically, classifier 106-n calculates the number E of extreme values. t For the characteristic data distribution 102T of n, the extreme values ​​belong to the number of extreme values ​​E respectively. s The probability of a category (the probability of belonging to a category) corresponding to the feature data distribution 102S for n features is given. Additionally, sometimes the learning model is simply represented as a model.

[0061] As a specific example, the model generation unit 16 uses KNN (k-nearest neighbor) or SVC (Support Vector Classification) as learning methods.

[0062] Classifier storage unit 17 stores classifiers 106-1 to 106-N; it is also referred to as the classifier 1 to N storage unit. The number of extreme values ​​E for classifier 106-n is also specified. s The characteristic data distribution of n corresponds to 102S.

[0063] Model Decision Unit 18 uses extreme coordinate group G t The classifier 106 used by the similarity calculation unit 19 is determined. As a specific example, the number of extreme values ​​E... t For m(=|{(x i t y i t In the case of )}|), the model determination unit 18 obtains the classifier 106-m from the classifier storage unit 17 and determines the obtained classifier 106-m as the classifier 106 used by the similarity calculation unit 19.

[0064] The model decision unit 18 generates inference data 108 containing the determined classifier 106. The inference data 108 contains the extreme value coordinate set G. t Sum and extreme value coordinate system G t The corresponding classifier is 106. The inference data 108-m is the number of extreme values ​​E. t The extreme coordinate system G of m t The inference data is 108.

[0065] Model decision unit 18 selects from at least one learned model, based on the extreme value coordinate set G t The number of extreme values ​​included will determine the learning model as the determining model.

[0066] The similarity calculation unit 19 uses the classifier 106 and the extreme value coordinate group G contained in the inference data 108. t Calculate the similarity between at least one feature data distribution 102S and feature data distribution 102T, and output similarity data 109 containing the calculated similarities. Similarity data 109 includes the data ID. s Tags and data IDs s The corresponding similarity.

[0067] The similarity calculation unit 19 uses the determination model and extreme value coordinate group G t Calculate the similarity. Here, the model is determined based on the extreme value coordinate set G. sTherefore, the similarity calculation unit 19 calculates the similarity based on the extreme value coordinates G corresponding to the feature data distribution 102S. s and extreme value coordinate system G t Find the similarity between feature data distribution 102S and feature data distribution 102T.

[0068] The following provides a specific example of similarity. [Equation 3] According to each estimated category c j Show each (x) i t y i t The output is the probability of belonging to a class of input. Here, the output is expressed in the form of "estimated class (probability of belonging to class)". Furthermore, the number of outputs for a single input is the number of extreme values ​​E. s Let m be the total number of characteristic data distributions 102S.

[0069] [Formula 3]

[0070]

[0071] As a specific example, this represents the relationship between the feature data distribution 102T and the estimated category c. j The similarity score (c j As shown in [Equation 4]. Additionally, score(c j This also represents the similarity between the destination data 3 and each candidate source data 2.

[0072] [Formula 4]

[0073]

[0074] Here, w i This indicates an emphasis on each extreme value (x) i s y i s The weighted parameter for the degree of ).

[0075] The data decision unit 20, also known as the migration source data decision unit, determines the migration source data from at least one migration source data candidate 2 based on the similarity data 109 and the decision conditions, and outputs decision data 4 containing the determined migration source data. Decision data 4 includes a label representing the migration source data candidate 2 and the similarity score corresponding to that candidate 2. The decision conditions are the criteria under which the data decision unit 20 determines the migration source data candidate 2 as the migration source data. Specifically, the decision conditions refer to a similarity score above a certain threshold. The data decision unit 20 determines the migration source data candidate 2 corresponding to the similarity score that satisfies the decision conditions as the migration source data.

[0076] As a specific example, the data decision unit 20 determines the data ID corresponding to the similarity that satisfies the decision conditions. s Output the determined data ID s The corresponding migration source data candidate 2 and similarity.

[0077] Figure 4 An example of the hardware structure of the similarity calculation device 1 according to this embodiment is shown. The similarity calculation device 1 is composed of a computer. The similarity calculation device 1 may also be composed of multiple computers.

[0078] Furthermore, the similarity calculation device 1 can operate in a server (computer) in cloud computing or in a server (computer) in edge computing.

[0079] As shown in the figure, the similarity calculation device 1 is a computer with hardware such as a processor 51, memory 52, auxiliary storage device 53, input / output IF (Interface) 54, and communication device 55. These hardware components are interconnected via signal lines 59.

[0080] Processor 51 is an IC (Integrated Circuit) that performs computational processing and controls the hardware of the computer. As a specific example, processor 51 is a CPU (Central Processing Unit), a DSP (Digital Signal Processor), or a GPU (Graphics Processing Unit).

[0081] The similarity calculation device 1 may also have multiple processors instead of processor 51. The multiple processors share the functions of processor 51.

[0082] Memory 52 is typically a volatile storage device. Memory 52 is also referred to as main storage or main memory. As a specific example, memory 52 is RAM (Random Access Memory). Data stored in memory 52 is saved to auxiliary storage device 53 as needed.

[0083] The auxiliary storage device 53 is typically a non-volatile storage device. As a specific example, the auxiliary storage device 53 is ROM (Read Only Memory), HDD (Hard Disk Drive), or flash memory. Data stored in the auxiliary storage device 53 is loaded into memory 52 as needed.

[0084] The memory 52 and the auxiliary storage device 53 can also be integrated into one unit.

[0085] Input / output IF 54 is a port for connecting input and output devices. As a specific example, input / output IF 54 is a USB (Universal Serial Bus) terminal. As a specific example, input devices are a keyboard and mouse. As a specific example, output devices are a monitor.

[0086] The communication device 55 is a receiver and a transmitter. As a specific example, the communication device 55 is a communication chip or a NIC (Network Interface Card).

[0087] Each part of the similarity calculation device 1 may also use a communication device 55 when communicating with other devices. Each part of the similarity calculation device 1 may receive data via an input / output IF 54, or via the communication device 55.

[0088] The auxiliary storage device 53 stores a similarity calculation program. The similarity calculation program is a program that enables the computer to perform the functions of each component of the similarity calculation device 1. The similarity calculation program can also consist of multiple files. The similarity calculation program is loaded into memory 52 and executed by the processor 51. The functions of each component of the similarity calculation device 1 are implemented through software.

[0089] The data used during the execution of the similarity calculation program and the data obtained by executing the similarity calculation program are appropriately stored in a storage device. Each part of the similarity calculation device 1 appropriately utilizes a storage device. As a specific example, the storage device consists of at least one of the following: main memory 52, auxiliary storage device 53, registers in processor 51, and cache memory in processor 51. Furthermore, data and information sometimes have the same meaning. The storage device may also be independent of the computer. The dataset storage unit 15 and the classifier storage unit 17 are each composed of a storage device.

[0090] The functions of memory 52 and auxiliary storage device 53 can also be achieved through other storage devices.

[0091] Similarity calculation programs can also be recorded on non-volatile recording media that can be read by a computer. Specific examples of non-volatile recording media include optical discs or flash memory. Similarity calculation programs can also be provided as program products.

[0092] ***Instructions for Action***

[0093] The operation of the similarity calculation device 1 is equivalent to a similarity calculation method. Furthermore, the program that implements the operation of the similarity calculation device 1 is equivalent to a similarity calculation program.

[0094] The similarity calculation device 1 evaluates the similarity between feature data distribution 102S and feature data distribution 102T based on the following two basic principles.

[0095] Policy (1): When a certain extreme value of feature data distribution 102S is close to a certain extreme value of feature data distribution 102T, the similarity calculation device 1 considers that feature data distribution 102S and feature data distribution 102T are partially similar at that certain extreme value of feature data distribution 102S. Furthermore, when a certain extreme value of feature data distribution 102T is closer to a certain extreme value of feature data distribution 102S, the similarity calculation device 1 may also consider that feature data distribution 102S and feature data distribution 102T are partially more similar.

[0096] Policy (2): The more feature data distributions 102S with partially similar extreme values, the higher the similarity calculation device 1 will calculate the similarity.

[0097] Figure 5 This is a diagram illustrating the policies (1) and (2) described above. The diagram shows two feature data distributions 102S (source data distribution D2 and source data distribution D3) and one destination data distribution (destination data distribution D1). Furthermore, circles represent maxima, quadrilaterals represent minima, and circles centered on the extreme values ​​of destination data distribution D1 represent regions close to each extreme value. These close regions are designated as neighborhood regions R1 to R4. Additionally, when an extreme value of the source data distribution is close to an extreme value of the destination data distribution D1, dashed lines connect the extreme values.

[0098] According to the figure, the extreme values ​​of the source data distribution D2 are close to the extreme values ​​of the destination data distribution D1 in the surrounding regions R1 to R4. On the other hand, the extreme values ​​of the source data distribution D3 are close to the extreme values ​​of the destination data distribution D1 only in the surrounding regions R1, R2, and R4. That is, the source data distribution D2 has four partially similar extreme values, and the source data distribution D3 has three partially similar extreme values. Therefore, the source data distribution D2 has more partially similar extreme values ​​than the source data distribution D3. Therefore, the similarity calculation device 1 calculates each similarity in a manner that the similarity corresponding to the source data distribution D2 is higher than the similarity corresponding to the source data distribution D3.

[0099] <Learning Phase>

[0100] Figure 6 This is a flowchart illustrating an example of the actions during the learning phase of the similarity calculation device 1. Using Figure 1 The diagram illustrates the learning phase.

[0101] (Step S101: Feature extraction processing)

[0102] The feature extraction unit 11 receives migration source data candidate 2 as input, uses the received migration source data candidate 2 to generate feature vector 101S, and outputs the generated feature vector 101S.

[0103] (Step S102: Data distribution calculation and processing)

[0104] The data distribution calculation unit 12 receives the feature vector 101S as input, uses the received feature vector 101S to generate the feature data distribution 102S, and outputs the generated feature data distribution 102S.

[0105] Figure 7 This is a diagram illustrating specific examples of steps S101 and S102. This diagram is used to illustrate specific examples of steps S101 and S102. Furthermore, in this diagram, (a) representing migration source data candidate 2 is shown in black, but it is actually data representing a color image of a tiger.

[0106] When the candidate source data 2 is an image and the pattern or design in the image has features, the feature extraction unit 11 outputs data representing the frequency of occurrence of pixel values ​​as a feature quantity. The data representing the frequency of occurrence can also be represented using a histogram. The feature extraction unit 11 sets the probability of the proportion of pixel value x in one image as a feature quantity. As a specific example, after converting the color image to grayscale, the feature extraction unit 11 calculates the probability mass function p(x) with pixel value x as a probability variable through kernel density estimation, and sets (x, p(x)) as the feature vector 101S. (b) shows the data for converting (a) to grayscale.

[0107] The data distribution calculation unit 12 receives (x, p(x)), performs a scaling transformation on (x, p(x)), and calculates the characteristic data distribution 102S. (c) shows the characteristic data distribution 102S calculated by the data distribution calculation unit 12.

[0108] Here, (x, p(x)) as feature vector 101S can be understood as feature data distribution 102S. Since feature vector 101S can be understood as feature data distribution 102S, the data distribution calculation unit 12 can also directly set the received feature vector 101S as feature data distribution 102S, and set the data obtained by scaling the received feature vector 101S as feature data distribution 102S. The purpose of scaling feature vector 101S is to facilitate comparison between feature data distributions 102S in subsequent processing. As a specific example of scaling, the data distribution calculation unit 12 scales the range of the data distribution from minimum to maximum value to the range [0, 1].

[0109] (Step S103: Extreme value calculation and processing)

[0110] The extreme value calculation unit 13 receives the feature data distribution 102S as input, generates extreme value data 103S using the received feature data distribution 102S, and outputs the generated extreme value data 103S. Specifically, the extreme value calculation unit 13 calculates the number of each maximum and minimum value and the extreme value coordinate set G representing the coordinates of each maximum and minimum value based on the feature data distribution 102S. s Output the calculated data as the extreme value data 103S.

[0111] (Step S104: Grouping Process)

[0112] The grouping unit 14 receives extreme value data 103S as input and saves the same number of learning data 104 as the number of extreme values ​​represented by the received extreme value data 103S to the dataset storage unit 15.

[0113] The similarity calculation device 1 repeats steps S101 to S104 in an amount corresponding to the number of prepared migration source data candidates 2.

[0114] (Step S105: Model Generation Processing)

[0115] The model generation unit 16 receives the dataset 105 as input, learns using the received dataset 105, generates a classifier 106, and saves the generated classifier 106 to the classifier storage unit 17.

[0116] Figure 8 This is a flowchart illustrating an example of the actions during the reasoning phase of the similarity calculation device 1. Using Figure 1 The diagram illustrates the reasoning stage.

[0117] <Reasoning Stage>

[0118] (Step S111: Feature Extraction Processing)

[0119] The feature extraction unit 11 receives the migration destination data 3 as input, uses the received migration destination data 3 to generate a feature vector 101T, and outputs the generated feature vector 101T.

[0120] (Step S112: Data distribution calculation and processing)

[0121] The data distribution calculation unit 12 receives the feature vector 101T as input, uses the received feature vector 101T to generate the feature data distribution 102T, and outputs the generated feature data distribution 102T.

[0122] (Step S113: Extreme value calculation and processing)

[0123] The extreme value calculation unit 13 receives the feature data distribution 102T, uses the received feature data distribution 102T to generate extreme value data 103T, and outputs the generated extreme value data 103T.

[0124] (Step S114: Model Decision Processing)

[0125] Hereinafter, we assume that the number of extrema is E. t Let's explain m.

[0126] The model decision unit 18 receives extreme value data 103T as input, retrieves the classifier 106 (i.e., classifier 106-m) corresponding to the received extreme value data 103T from the classifier storage unit 17, and uses the extreme value coordinate group G contained in the received extreme value data 103T. t Generate inference data 108-m from the obtained classifier 106-m, and output the generated inference data 108-m.

[0127] (Step S115: Similarity Calculation Processing)

[0128] The similarity calculation unit 19 receives inference data 108-m as input, and uses the classifier 106-m and extreme value coordinate group G contained in the received inference data 108-m. t Calculate the similarity score and output the similarity data as 109.

[0129] As a specific example, firstly, the similarity calculation unit 19 receives the extreme value coordinate group G. t Extract the extreme values ​​from the data. At this point, the similarity calculation unit 19 generates a total of m sets (x, y, z) of extreme value coordinates. i t y i t (i = 1, ..., m). Additionally, the extreme values ​​are also test data.

[0130] Next, the similarity calculation unit 19 inputs m test data one by one into the classifier 106-m, and calculates the probability of belonging to each estimated category. The estimated category is the category to which the test data is estimated to belong. The similarity calculation unit 19 estimates that each test data belongs to each of the candidate categories, and calculates the probability of each test data belonging to each category. Hereinafter, (x i t y i t Let C be the set of labels for the estimated categories. t , will C t The element is set to c j .

[0131] Next, regarding Calculate score(c) j Output data ID s Tags and data IDs s The corresponding similarity groups (c j score(c j )).

[0132] (Step S116: Data Decision Processing)

[0133] The data decision unit 20 receives similarity data 109 as input, uses the received similarity data 109 to generate decision data 4, and outputs the generated decision data 4.

[0134] Figure 9 This diagram illustrates a specific example of how the data determination unit 20 determines migration source data candidate 2. In this diagram, the migration destination data 3 is an image of a tiger, and the labels and similarities of the migration source data candidate 2 corresponding to each image of migration source data candidate 2 are recorded in the form of "labels (similarity) of migration source data candidate 2". Furthermore, each image is actually a color photograph. In this example, the data determination unit 20 sets 0.5 as the threshold and sets a similarity greater than the threshold as the determination condition. At this time, the determined data 4 are (leopard1, 0.972), (cat1, 0.891), (tiger1, 0.779), (tiger2, 0.697), (cheetah2, 0.594), and (cat2, 0.567). The similarity of other images is below the threshold; therefore, the data determination unit 20 does not select other images.

[0135] As described above, the similarity calculation device 1 learns from one or more migration source data candidates 2 to generate a classifier 106, and then uses the generated classifier 106 to perform inference based on one migration destination data 3. Based on the inference result, it outputs the migration source data and the similarity to the migration source data.

[0136] ***Explanation of the effects of Implementation Method 1***

[0137] As described above, as a specific example, firstly, the similarity calculation device 1 of this embodiment outputs a data distribution calculation unit 12 that calculates (x, p(x)) based on the feature vector 101S, as a two-dimensional feature data distribution 102S. Next, the extremum calculation unit 13 calculates the minimum and maximum values ​​of the feature data distribution 102S. Then, the extremum coordinate set G is determined. s The extreme values ​​and extreme value coordinates G t To what extent are the extreme values ​​of the given values ​​close to each other?

[0138] Therefore, according to this embodiment, compared to directly using feature vector 101, the dimensionality of the information representing the features can be reduced to use feature data distribution 102. Furthermore, the computational load for similarity is reduced, narrowing the computational object down to the extreme value of the features with the migration destination. Therefore, according to this embodiment, a similarity close to that obtained using feature data distribution 102 itself can be calculated in a shorter processing time.

[0139] Furthermore, the similarity calculation device 1 of this embodiment calculates the probability of belonging to a category corresponding to the migration source data candidate 2 according to each extreme value of the feature data distribution 102T, and multiplies each category probability by a weighting parameter w. i The similarity is then calculated by combining these factors. Therefore, according to this embodiment, by targeting specific extreme values ​​of w for the feature data distribution 102T... i Set relatively high (as a specific example, w) i >1), can increase the influence of specific extreme values, and in addition, can increase w i Set it to a low level (for example, w) i <1), in order to reduce the influence of specific extreme values.

[0140] Therefore, according to this embodiment, in the calculation of similarity, it is possible to calculate the similarity of the extreme values ​​(features) specified by the user.

[0141] In the technology disclosed in Patent Document 1, when evaluating image similarity, using existing similarity calculation methods, users need to select an appropriate method that takes into account detection expectations such as processing load. Furthermore, according to this technology, it is difficult to control the degree of similarity. Specifically, according to this technology, it is impossible to implement processing that evaluates multiple types of similarity using a single evaluation function, such as fingerprint authentication which performs near-perfect similarity considering only environmental noise, or near-fuzzy retrieval that ignores individual differences and labels, as in image segmentation or behavior recognition.

[0142] However, as described above, the similarity calculation device 1 according to this embodiment has a low computational load related to similarity, and it is possible to evaluate the similarity of multiple categories by flexibly utilizing weights.

[0143] ***Other Structures***

[0144] <Variation Example 1>

[0145] As for (x, p(x)) as the feature data distribution 102, it can be arbitrary as long as it fully reflects the characteristics of the migration source data candidate 2 and the migration destination data 3 and approximates the wave shape.

[0146] As a specific example, p(x) uses the BoF (Bag-of-Features) method shown in Non-Patent Document 1. BoF is a method that clusters feature vectors extracted from metadata such as images, generating category IDs and frequency (histograms) related to the clustering. As a specific example of using BoF, when receiving an image, the feature extraction unit 11 outputs either HOG (Histograms of Oriented Gradients) as shown in Non-Patent Document 2 or SIFT (Scaled Invariance Feature Transform) as shown in Non-Patent Document 3, as a feature vector 101. Then, the data distribution calculation unit 12 uses BoF to assign a category ID to x, assigns the frequency of occurrence of the category corresponding to that category ID to p(x), and outputs (x, p(x)) as the feature data distribution 102.

[0147] [Non-patent literature 1]

[0148] Csurka, G., Dance, CR, Fan, L., Willamowski, J. and Bray, C.: Visual Categorization with Bags of Keypoints, ECCV (European Conference on Computer Vision) International Workshop on Statistical Learning in Computer Vision, pp. 1-22 (2004).

[0149] [Non-patent literature 2]

[0150] Dalal, N. and Triggs, B.: Histograms of Oriented Gradients for Human Detection, 2005 IEEE (Institute of Electrical and Electronics Engineers) Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), vol.1, pp.886-893, doi:10.1109 / CVPR.2005.177(2005).

[0151] [Non-patent literature 3]

[0152] Lowe, DG: Distinctive Image Features from Scale-Invariant Keypoints, Int.J.Comput.Vision, Vol.60, No.2, pp.91-110 (2004).

[0153] <Variation Example 2>

[0154] In the case where the migration source data candidate 2 is multidimensional time series data obtained from sensors, etc., as a specific example, the feature extraction unit 11 inputs multiple sampled data in time series order, and sets the value obtained by summarizing each sampled data through principal component analysis (reducing dimensionality) as the feature vector 101S.

[0155] In this modified example, the data distribution calculation unit 12 can also generate a feature data distribution 102S by combining the feature vectors 101S in time series order. Furthermore, it is assumed that this feature data distribution 102S is a data distribution capable of calculating extrema.

[0156] <Variation Example 3>

[0157] Figure 10 This section shows an example of the hardware structure of the similarity calculation device 1 in this modified example.

[0158] As shown in the figure, the similarity calculation device 1 has a processing circuit 58 in place of at least one of the processor 51, memory 52 and auxiliary storage device 53.

[0159] The processing circuit 58 is hardware that implements at least a portion of the components of the similarity calculation device 1.

[0160] The processing circuit 58 can be dedicated hardware, or it can be a processor that executes the program stored in memory 52.

[0161] In the case where the processing circuit 58 is dedicated hardware, as a specific example, the processing circuit 58 is a single circuit, a composite circuit, a programmable processor, a parallel programmable processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0162] The similarity calculation device 1 may also have multiple processing circuits that replace the processing circuit 58. The multiple processing circuits share the functions of the processing circuit 58.

[0163] In the similarity calculation device 1, some functions can be implemented through dedicated hardware, while the remaining functions can be implemented through software or firmware.

[0164] As a specific example, the processing circuit 58 is implemented by hardware, software, firmware, or a combination thereof.

[0165] The processor 51, memory 52, auxiliary storage device 53, and processing circuit 58 are collectively referred to as the "processing circuit". That is, the functions of each functional structural element of the similarity calculation device 1 are realized through the processing circuit.

[0166] The similarity calculation device 1 in other embodiments may also have the same structure as this modified example.

[0167] Implementation Method 2

[0168] Hereinafter, with reference to the accompanying drawings, we will mainly describe the aspects that differ from the embodiments described above.

[0169] This embodiment provides an overview. When generating a learning model, the similarity calculation apparatus 1 of this embodiment transforms the coordinate system representing extreme values ​​to prevent erroneous mapping between the minimum values ​​of feature data distribution 102S and the maximum values ​​of feature data distribution 102T, or between the maximum values ​​of feature data distribution 102S and the minimum values ​​of feature data distribution 102T. As a result, the learning model can easily identify maximum and minimum values.

[0170] Figure 11The diagram illustrates a specific example of an image transformed by a coordinate system representing extrema. Figure (a) shows the image before the coordinate system transformation, and (b) shows the image after the coordinate system of the extrema is transformed by projecting the extrema shown in (a) onto a higher-dimensional space. Furthermore, the dimension of this higher-dimensional space is not limited to 3. The maxima set in (b) represents the set of points projected with the maxima shown in (a), and the minima set represents the set of points projected with the minima shown in (a). The mismapping in (a) indicates that a minima in the destination data distribution D1 is close to a maxima in the source data distribution D3. The minima and maxima associated with the mismapping are explicitly separated in (b) by a separating hyperplane.

[0171] ***Structure Description***

[0172] Figure 12 The diagram shows a structural example of the similarity calculation device 1 according to this embodiment. As shown in the figure, in addition to the structural elements of the similarity calculation device 1 of Embodiment 1, the similarity calculation device 1 also has a coordinate transformation unit 21.

[0173] The coordinate transformation unit 21 uses the extreme value data 103S to generate the transformed coordinate set C. s The output contains the generated transformed coordinate set C. s The transformed extreme value data 107S. The transformed extreme value data 107S contains the transformed coordinate group C. s Data ID s And the number of extreme values ​​E s This is also known as the migration source transformation coordinate group. Transformation coordinate group C s It is based on the extreme value coordinate system G s The set of coordinates obtained by transforming the coordinate systems of the included coordinates. Extreme value coordinate set G s The included coordinates are also called the original source extrema before transformation. The coordinate transformation unit 21 calculates the coordinates obtained by projecting the original source extrema before transformation onto a space with a higher dimension than the original source extrema before transformation, and uses these coordinates as the source extrema. Transformed coordinate group C s It is also the extreme value coordinate system G s The subordinate concept. Transformation coordinate group C s The coordinates included are also the extreme values ​​of the migration source.

[0174] Furthermore, the coordinate transformation unit 21 uses the extreme value data 103T to generate the transformed coordinate set C. t The output contains the transformed coordinate system C. t The transformed extreme value data 107T. The transformed extreme value data 107T contains the transformed coordinate group C. t And the number of extreme values ​​E t This is also known as the destination transformation coordinate group. Transformation coordinate group C tIt is based on the extreme value coordinate system G t The set of coordinates obtained by transforming the coordinate systems of the included coordinates. Extreme value coordinate set G t The included coordinates are also referred to as the original destination extrema. The coordinate transformation unit 21 calculates the coordinates obtained by projecting the original destination extrema onto a space with a higher dimension than the original destination extrema, and uses these coordinates as the destination extrema. Transformed coordinate group C t It is also the extreme value coordinate system G t The subordinate concept. Transformation coordinate group C t The included coordinates are also the extreme values ​​of the migration destination.

[0175] Transformation coordinate group C s The coordinates included are defined as shown in [Equation 5]. [Equation 5] is also the transformed coordinate group C. s A specific example. In this example, the coordinate transformation unit 21 transforms the extreme coordinate group G... s Transform all the maxima contained in the extreme value set G to 1, and transform the extreme value coordinate set G. s Each local minimum value is transformed to -1, and the transformed coordinates are appended to the extreme value coordinate set G. s From the included coordinates, we obtain the transformed coordinate set C. s Furthermore, in this example, the coordinate group C is transformed. s The extreme values ​​included are represented as (x i s y i s , z i Transform coordinate system C t The coordinates included are defined in the same way as in [Equation 5].

[0176] [Formula 5]

[0177]

[0178] ***Instructions for Action***

[0179] The main focus is on explaining the differences in operation between the similarity calculation device 1 and the one described in Implementation Method 1.

[0180] <Learning Phase>

[0181] Figure 13 This is a flowchart illustrating an example of the actions during the learning phase of the similarity calculation device 1. Using Figure 12 The diagram illustrates the learning phase.

[0182] (Step S201: Coordinate Transformation Process)

[0183] The coordinate transformation unit 21 receives extreme value data 103S as input, uses the received extreme value data 103S to generate transformed extreme value data 107S, and outputs the generated transformed extreme value data 107S.

[0184] (Step S104: Grouping Process)

[0185] This step involves replacing the extreme value data 103S in the grouping process related to the learning phase in Implementation 1 with the transformed extreme value data 107S.

[0186] <Reasoning Stage>

[0187] Figure 14 This is a flowchart illustrating an example of the actions during the reasoning phase of the similarity calculation device 1. Using Figure 12 The diagram illustrates the reasoning stage.

[0188] (Step S211: Coordinate Transformation Process)

[0189] The coordinate transformation unit 21 receives extreme value data 103T as input, uses the received extreme value data 103T to generate transformed extreme value data 107T, and outputs the generated transformed extreme value data 107T.

[0190] (Step S114: Model Decision Processing)

[0191] This step involves replacing the extreme value data 103T in the model determination process related to the reasoning stage in Implementation 1 with the transformed extreme value data 107T.

[0192] ***Explanation of the effects of Implementation Method 2***

[0193] As described above, as a specific example, the similarity calculation device 1 of this embodiment transforms the coordinate system of each extreme value by introducing a parameter z representing whether each extreme value is a maximum or a minimum value for each extreme value of the feature data distribution 102, thereby separating the distance between the set of maximum values ​​and the set of minimum values. Therefore, according to the similarity calculation device 1 of this embodiment, the risk of incorrect mapping between the maximum value of the migration destination and the minimum value of the migration source, or between the minimum value of the migration destination and the maximum value of the migration source, is reduced.

[0194] Therefore, the similarity calculation device 1 according to this embodiment can obtain a higher accuracy similarity compared to the similarity calculation device 1 of embodiment 1.

[0195] ***Other Structures***

[0196] <Variation Example 4>

[0197] As shown in [Equation 6], the coordinate system C is transformed. sThe coordinates can also be relative to x. i s and y i s Multiply by z respectively i And the form obtained.

[0198] [Formula 6]

[0199]

[0200] Even when the coordinate system is transformed as in [Equation 5], the maximum and minimum values ​​can be guaranteed to be in the three-dimensional Euclidean space coordinate system (R). 3 The Euclidean distance in the configuration is at least 2.

[0201] In this variation, x and y are multiplied by z respectively to easily configure the maxima and minima with a larger distance after coordinate system transformation. The maxima (x) are already assigned... i s y i s ) and minimum value (x) j s y j s In the case of ), the distance between the two after coordinate system transformation is shown in [Equation 7].

[0202] [Formula 7]

[0203]

[0204] Consider [Equation 7]. When the variable x of feature data distribution 102 is normalized through [0, 1], x ≧ 0. Since y is a histogram, y ≧ 0. Therefore, x i s -x j s <x i s +x j s and y i s -y j s <y i s +y j s This holds true. Therefore, as long as the maxima and minima are slightly separated, it can be guaranteed that the transformed maxima and minima are at least easier to separate than 2.

[0205] Therefore, when using algorithms like SVM (Support Vector Machine) in inference, it is easy to use the separating hyperplane to identify maxima and minima.

[0206] Therefore, according to this modified example, it is possible to reduce the probability of incorrectly inferring the minimum value of the feature data distribution 102S when the maximum value of the feature data distribution 102T is input, and the probability of incorrectly inferring the maximum value of the feature data distribution 102S when the minimum value of the feature data distribution 102T is input.

[0207] Implementation Method 3

[0208] Hereinafter, with reference to the accompanying drawings, we will mainly describe the aspects that differ from the embodiments described above.

[0209] The main difference between this embodiment and Embodiment 1 is that the model generation unit 16 uses a dataset 105 corresponding to the number of extreme values ​​within the allowable range to expand the range of migration source data candidates 2 that are the objects of similarity calculation. This allowable range is also called the dataset allowable range. The dataset allowable range represents the range around the number of migration source extreme values ​​corresponding to the dataset 105.

[0210] Figure 15 This diagram illustrates a specific example where, during model generation, the model generation unit 16 selects images of dataset 105 corresponding to the number of extreme values ​​within the allowable range. The model generation unit 16 selects dataset 105 based on the number of extreme values ​​corresponding to dataset 105. Furthermore, for ease of explanation, this diagram illustrates the case where the learning model is determined based on feature data distribution 102; however, in reality, the learning model is determined based on the extreme value data 103T corresponding to feature data distribution 102T and dataset 105 corresponding to feature data distribution 102S.

[0211] In this figure, the number of extreme values ​​in the destination data distribution D1 is 4, the number of extreme values ​​in the source data distribution D2 is 3, and the number of extreme values ​​in the source data distribution D3 is 5. (a) shows an image of the dataset 105 selected by the model generation unit 16 without considering the allowable range. In (a), there is no source data distribution with the same number of extreme values ​​as the destination data distribution D1, therefore, the model generation unit 16 cannot select dataset 105. On the other hand, (b) shows an image of the dataset 105 selected by the model generation unit 16 considering the allowable range. The model generation unit 16 allows the number of extreme values ​​within a range of ±1. The number of extreme values ​​in both the source data distribution D2 and the source data distribution D3 is within a range of ±1 of the number of extreme values ​​in the destination data distribution D1, therefore, the model generation unit 16 selects dataset 105 corresponding to the source data distributions D2 and D3, respectively.

[0212] ***Structure Description***

[0213] Figure 16 This section illustrates a structural example of the similarity calculation apparatus 1 according to this embodiment. The main difference between the similarity calculation apparatus 1 of this embodiment and the similarity calculation apparatus 1 of Embodiment 1 is that the model generation unit 16 receives a dataset group 111 instead of a dataset 105. The dataset group 111 is a collection composed of multiple datasets 105. The model generation unit 16 may also receive multiple datasets 105 and generate the dataset group 111 using the received multiple datasets 105.

[0214] As a specific example, dataset group 111 is dataset group 111-(n±b). Dataset group 111-(n±b) is a set consisting of datasets 105, each corresponding to the number of extreme values ​​in the range from nb to n+b. Here, b is a parameter corresponding to the allowable range, b≧0 and b∈Z. Alternatively, there may not be a dataset 105 corresponding to the number of any more than one extreme value in the range from nb to n+b.

[0215] Equation 8 shows a specific example of the dataset group 111-(n±b). The meanings of the labels are the same as in Equation 2. Furthermore, the dataset group 111-(n±b) is represented as a set g. n±b .

[0216] Dataset group 111 contains migration source extrema corresponding to the number of extrema within the permissible range of the dataset, and is also a subordinate concept of dataset 105. In dataset group 111-(n±b), the number of migration source extrema corresponding to dataset 105 is n, and the permissible range of the dataset is from nb to n+b.

[0217] [Formula 8]

[0218]

[0219] The model generation unit 16 obtains dataset 105-n from dataset storage unit 15 to replace dataset group 111, and generates classifier 106 using the obtained dataset group 111. As a specific example, the model generation unit 16 generates classifier 106-n using dataset group 111-(n±b).

[0220] ***Instructions for Action***

[0221] The flowchart illustrating the operation of the similarity calculation device 1 in this embodiment is the same as the flowchart illustrating the operation of the similarity calculation device 1 in Embodiment 1. Hereinafter, the differences in operation between the similarity calculation device 1 and that in Embodiment 1 will be primarily explained.

[0222] <Learning Phase>

[0223] (Step S105: Model Generation Processing)

[0224] This process replaces dataset 105 in the model generation process related to the learning phase in Implementation 1 with dataset group 111-(n±b). That is, the model generation unit 16 uses dataset group 111-(n±b) when generating classifier 106-n.

[0225] The following processing is the same as in Implementation Method 1. In addition, the model generation unit 16 saves the learning model learned using the dataset group 111-(n±b) as classifier 106-n to the classifier storage unit 17.

[0226] Furthermore, when the training data is set as dataset group 111-(n±b), the same extreme value coordinate group G is used in the learning process from classifier 106-(nb) to classifier 106-(n+b). s {(x i s1 y i s1 )}(where, |(x i s1 y i s1 )|=n、i=1,…,n、s1∈S). Therefore, in {(x i s1 y i s1 )} and {(x i t y i t When the similarity between {(x)} is high, using {(x)} i s1 yi s1 In any classifier 106 generated by )}, those belonging to {(x)} i t y i t The probability of belonging to the category corresponding to s1 increases. As a result, when the category corresponding to s1 is set to c1, the value of the similarity score(c1) increases. Therefore, it can be considered that even if an allowable range is set for the number of extreme values ​​of the benchmark training data used as the learning of the selected learning model, it is possible to appropriately classify the extreme values ​​of the transfer destination data 3.

[0227] ***Explanation of the effects of Implementation Method 3***

[0228] As described above, the similarity calculation apparatus 1 of this embodiment sets a data set allowable range for the number of extreme values ​​of the benchmark training data used in the learning of the selected learning model for calculating similarity. Specifically, the similarity calculation apparatus 1 expands the candidate range of migration source data candidates 2, which are the objects of similarity calculation, by importing parameter b.

[0229] Therefore, when the number of extreme values ​​is different but the feature data distribution 102S and the feature data distribution 102T are generally similar, the similarity calculation device 1 can determine the similarity between the two data distributions.

[0230] Figure 17 This diagram illustrates the operation of the similarity calculation device 1 when feature data distributions 102S and 102T are generally similar. In this diagram, the number of extreme values ​​for the destination data distribution D1 and the source data distribution D2 differs, but regardless of the value of x, the corresponding y values ​​are close between the two data distributions. That is, the two data distributions are generally similar. Furthermore, the nearby regions R5 and R6 are regions where the extreme values ​​of the destination data distribution D1 are close to those of the source data distribution D2.

[0231] The more extreme values ​​of the source data distribution D2 that are near the extreme values ​​of the destination data distribution D1, the higher the probability that the similarity calculation device 1 will classify the category corresponding to the destination data distribution D1 into the category corresponding to the source data distribution D2.

[0232] Furthermore, the similarity calculation device 1 can control the degree of similarity using parameter b. Specifically, the smaller the value of parameter b, the closer the similarity calculation device 1 is to a completely identical search; the larger the value of parameter b, the closer the similarity calculation device 1 is to a fuzzy search.

[0233] ***Other Structures***

[0234] <Variation Example 5>

[0235] The allowed range of the dataset may also lack symmetry. As a specific example, the model generation unit 16 uses dataset set 111-k (k = n, n+1, ..., n+b) instead of dataset set 111-(n±b) as dataset set 111 used when generating classifier 106-n. Dataset set 111-k (k = n, n+1, ..., n+b) is dataset 105 corresponding to the number of extreme values ​​in the range from n to n+b.

[0236] <Variation Example 6>

[0237] The allowable range of the dataset can also be set dynamically.

[0238] As a specific example of this variation, if only datasets 105-1, 105-3, and 105-4 are obtained as dataset 105, consider selecting each dataset 105 and one dataset 105 that is closest to each dataset 105 on both sides, to generate dataset group 111 corresponding to each dataset 105. In this case, when using dataset 105-3 as the reference, datasets 105-1 and 105-4 are selected as the closest datasets 105. Thus, the model generation unit 16 can also consider whether there exists a dataset 105 corresponding to the number of each extreme value, select dataset 105, and generate dataset group 111.

[0239] Implementation Method 4

[0240] Hereinafter, with reference to the accompanying drawings, we will mainly describe the aspects that differ from the embodiments described above.

[0241] The main difference between this embodiment and Embodiment 1 is that the model determination unit 18 uses classifiers 106 corresponding to the number of extreme values ​​within the allowable range to expand the range of migration source data candidates 2 that are the objects of similarity calculation. This allowable range is also called the model allowable range. The model allowable range represents the range around the number of migration destination extreme values ​​contained in the migration destination extreme value group.

[0242] Figure 18 This shows a specific example of an image of the classifier 106 selected by the model generation unit 16, corresponding to the number of extreme values ​​within the allowable range. The method of observing this image is similar to... Figure 15 The observation method is the same. The model generation unit 16 selects at least one classifier 106 based on the number of extreme values.

[0243] In this figure, (a) shows an image where the model determination unit 18 selects classifier 106 without considering the allowable range. In (a), there is no classifier 106 corresponding to the number of extreme values ​​as the number of extreme values ​​in the migration destination data distribution D1; therefore, the model determination unit 18 cannot select classifier 106. On the other hand, (b) shows an image where the model determination unit 18 selects classifier 106 considering the allowable range. In (b), the allowable range is 3 or more and 5 or less, and the number of extreme values ​​in the migration source data distributions D2 and D3 is within the allowable range. Therefore, the model generation unit 16 selects classifiers 106-3 and 106-5, generating a classifier group 112 consisting of classifiers 106-3, 106-4, and 106-5, as the classifier group 112 corresponding to the feature data distribution 102-3.

[0244] ***Structure Description***

[0245] Figure 19 The structure of the similarity calculation device 1 of this embodiment is shown as an example. The main difference between the similarity calculation device 1 of this embodiment and the similarity calculation device 1 of embodiment 1 is that the model determination unit 18 outputs classifier group 112 instead of classifier 106-m.

[0246] As a specific example, classifier set 112 is classifier set 112-(m±d). Here, d is a parameter corresponding to the allowable range, d≧0 and d∈Z. Classifier set 112-(m±d) is a set of classifiers 106 from classifier 106-(md) to classifier 106-(m+d). Alternatively, there may not be a classifier 106 corresponding to the number of any more than one extreme value in the range from md to m+d.

[0247] Model Determination Unit 18 Generation and Extreme Value Coordinate Set G t The corresponding classifier group 112 is used to replace the extreme value coordinate group G. t The corresponding classifier 106 generates a set G containing extreme value coordinates. t Sum and extreme value coordinate system G t The data from the corresponding classifier group 112 is used as inference data 108.

[0248] The model determination unit 18 determines the decision model from the learning model group consisting of learning models that correspond to the number of migration source extrema within the model's tolerance range.

[0249] The similarity calculation unit 19 uses each classifier 106 contained in the classifier group 112 to replace the classifier 106, and calculates the similarity between at least one feature data distribution 102S and the feature data distribution 102T.

[0250] ***Instructions for Action***

[0251] The flowchart illustrating the operation of the similarity calculation device 1 in this embodiment is the same as the flowchart illustrating the operation of the similarity calculation device 1 in Embodiment 1. Hereinafter, the differences in operation between the similarity calculation device 1 and that in Embodiment 1 will be primarily explained.

[0252] <Reasoning Stage>

[0253] (Step S114: Model Decision Processing)

[0254] This process replaces the classifier 106-m in the model determination process related to the inference stage in Implementation 1 with the classifier group 112-(m±d). That is, the model determination unit 18 determines the classifier group 112-(m±d) to replace the classifier 106-m, and outputs the determined classifier group 112-(m±d) and the extreme value coordinate group G. t The inference data is 108.

[0255] (Step S115: Similarity Calculation Processing)

[0256] The similarity calculation unit 19 uses classifier group 112-(m±d) instead of classifier 106-m to calculate similarity.

[0257] The following describes in detail the similarity calculation process when this embodiment and embodiment 3 are combined. The similarity calculation unit 19 considers the possibility that the probability of category affiliation overlaps in the classifier 106, and modifies the calculation of the similarity score(c). j The method is as shown in [Equation 9], for the feature data distribution 102T with label c using classifier 106-k (k = md, m-d+1, ..., m+d). j Similarity sub-score k (c j The definitions are as follows. The explanations of the labels in [Equation 9] are the same as those in [Equation 4].

[0258] [Formula 9]

[0259]

[0260] When multiple classifiers 106-k can classify the same label, the similarity calculation unit 19 pays more attention to the number of sub-scores whose values ​​of k are closer to the extreme values ​​corresponding to the label for each of the multiple classifiers 106-k. That is, the more the sub-scores whose values ​​of k are closer to the extreme values, the more the similarity calculation unit 19 calculates the similarity.

[0261] Equation 10 shows the similarity score (c j) A specific example. Here, in the upper right section, k is the value that can be categorized into label c. j In classifier 106, the variable that the similarity calculation unit 19 prioritizes most is the classifier 106. k' represents a value other than k, which determines whether a classifier can be classified as label c. j The variables of classifier 106. Besides classifier 106-k, there are multiple other classes that can classify to label c. j In the case of classifier 106, k' consists of multiple values. As a specific example, fp(d) is a function that returns the maximum value that is inversely proportional to d when d = 0. As a specific example, fp(d) = 1 / (1+d) 2 Here, d represents the expression derived from the label c. j The variable is the distance from the number of corresponding extrema to each element of k'. Additionally, when k' consists of multiple values, prepare fp(d) corresponding to each element of k'.

[0262] As a specific example, consider the tag c j The number of corresponding extreme values ​​is m, and any one of the classifiers 106-k (k = md, m-d+1, ..., m+d) can be classified to label c. j In this case, k = m, k' = {md, ..., m-1, m+1, ..., m+d}. That is, score k (c j The similarity score is calculated using a classifier of 106-m. k’ (c j The value is obtained using similarity, which is calculated using classifier 106 other than classifier 106-m.

[0263] [Formula 10]

[0264]

[0265] ***Explanation of the effects of implementation method 4***

[0266] As described above, the similarity calculation apparatus 1 of this embodiment sets a model allowable range for the number of extreme values ​​that serve as the benchmark for selecting the learning model used to calculate similarity. Specifically, the similarity calculation apparatus 1 expands the range of candidate migration source data 2, which are the objects of similarity calculation, by importing parameter d.

[0267] Therefore, similar to the similarity calculation device 1 in Embodiment 3, in such Figure 17 In the case shown, the similarity calculation device 1 can determine the similarity between the two data distributions.

[0268] Furthermore, the similarity calculation device 1 can control the degree of similarity using parameter d. Specifically, the smaller the value of parameter d, the closer the similarity calculation device 1 is to a completely identical search; the larger the value of parameter d, the closer the similarity calculation device 1 is to a fuzzy search.

[0269] ***Other Structures***

[0270] <Variation Example 7>

[0271] The model's allowable range may also lack symmetry. As a specific example, the model decision unit 18 may also determine a classifier set 112-k (k = m, m+1, ..., m+d) instead of a classifier set 112-(m±d), as the classifier set 112 corresponding to the feature data distribution 102T with the number of extreme values ​​m. The classifier set 112-k (k = m, m+1, ..., m+d) is a classifier set 112 composed of classifiers 106 corresponding to the number of extreme values ​​in the range from m to m+d.

[0272] <Variation Example 8>

[0273] The model's tolerance range can also be set dynamically.

[0274] As a specific example of this variation, consider the following situation: When only datasets 105-1, 105-3, and 105-4 are obtained as dataset 105, classifiers 106 corresponding to each dataset 105 and classifiers 106 corresponding to the two datasets 105 closest to each other on either side of each dataset 105 are selected as classifier group 112. In this case, when the number of extreme values ​​is based on 3, classifiers 106-1, 106-3, and 106-4 are selected. Thus, the model determination unit 18 can also consider whether there exists a classifier 106 corresponding to the number of each extreme value and select classifier 106.

[0275] ***Other Implementation Methods***

[0276] It can realize the free combination of the above-described embodiments, or the modification of any structural elements of each embodiment, or the omission of any structural elements in each embodiment.

[0277] Furthermore, the implementation methods are not limited to those shown in Implementation Methods 1 to 4, and various modifications can be made as needed. The processes described using flowcharts and the like can also be appropriately modified.

[0278] Label Explanation

[0279] 1: Similarity calculation unit; 2: Candidate source data for migration; 3: Destination data for migration; 4: Decision data; 11: Feature extraction unit; 12: Data distribution calculation unit; 13: Extreme value calculation unit; 14: Grouping unit; 15: Dataset storage unit; 16: Model generation unit; 17: Classifier storage unit; 18: Model decision unit; 19: Similarity calculation unit; 20: Data decision unit; 21: Coordinate transformation unit; 51: Processor; 52: Memory; 53: Auxiliary storage device; 54: Input / output IF; 55: Communication device; 58: Processing circuit; 5 9: Signal line; 101, 101S, 101T: Feature vectors; 102, 102S, 102T: Feature data distribution; 103S, 103T: Extreme value data; 104: Learning data; 105: Data set; 106: Classifier; 107S, 107T: Transformed extreme value data; 108: Inference data; 109: Similarity data; 111: Data set group; 112: Classifier group; D1: Migration destination data distribution; D2, D3: Migration source data distribution; R1, R2, R3, R4, R5, R6: Neighborhood region.

Claims

1. A similarity calculation apparatus, wherein the similarity calculation apparatus performs transfer learning using candidate transfer source data and candidate transfer destination data, wherein the candidate transfer source data and the candidate transfer destination data are training data for learning from Internet of Things (IoT) devices collected in different environments, and wherein the candidate transfer source data and the candidate transfer destination data are image data, wherein, The similarity calculation device includes a similarity calculation unit that calculates the similarity between the source data distribution and the destination data distribution based on a set of migration source extreme values, including migration source extreme values ​​representing extreme values ​​in the source data distribution, and a set of destination extreme values, including migration destination extreme values ​​representing extreme values ​​in the destination data distribution. The source data distribution represents the distribution of feature quantities of the candidate source data, and the destination data distribution represents the distribution of feature quantities of the destination data. The extreme values ​​include maxima and minima. There is at least one of each of the candidate migration source data, the migration source data distribution, and the migration source extreme value group. Each candidate migration source data corresponds one-to-one with each migration source data distribution, and each migration source data distribution corresponds one-to-one with each migration source extreme value group. The similarity calculation unit calculates the similarity between each of the migration source data distributions and the migration destination data distributions based on the migration source extreme value group and the migration destination extreme value group corresponding to each of the migration source data distributions. Each of the migration source extrema is associated with an identifier that identifies a migration source data candidate corresponding to the migration source data distribution. Based on the number of migration source extrema contained in the migration source extrema group, each of the migration source extrema belongs to at least one of the datasets prepared for each value represented by the number of extrema, wherein the migration source data distribution has extrema corresponding to each of the migration source extrema. The similarity calculation device also has: A model generation unit generates a learning model corresponding to each of the datasets, the learning model being a model for estimating identifiers corresponding to assigned coordinate sets; and The model decision unit determines the learning model as the decision model based on the number of migration destination extrema contained in the migration destination extrema group. The similarity calculation unit uses the decision model and the extreme value set of the migration destination to calculate the similarity.

2. The similarity calculation device according to claim 1, wherein, Each of the datasets contains a number of migration source extrema corresponding to the migration source extrema associated with each of the datasets.

3. The similarity calculation device according to claim 1, wherein, Each of the datasets is a group of datasets containing migration source extrema corresponding to the number of extrema within the dataset tolerance range, wherein the dataset tolerance range represents the range around the number of migration source extrema corresponding to each of the datasets.

4. The similarity calculation device according to claim 1, wherein, The model decision unit determines the decision model as a set of learning models consisting of learning models such that each learning model corresponds to the number of migration source extrema within the model tolerance range, whereby the model tolerance range represents the range around the number of migration destination extrema contained in the migration destination extrema set.

5. The similarity calculation device according to claim 2, wherein, The model decision unit determines the decision model as a set of learning models consisting of learning models such that each learning model corresponds to the number of migration source extrema within the model tolerance range, whereby the model tolerance range represents the range around the number of migration destination extrema contained in the migration destination extrema set.

6. The similarity calculation device according to claim 3, wherein, The model decision unit determines the decision model as a set of learning models consisting of learning models such that each learning model corresponds to the number of migration source extrema within the model tolerance range, whereby the model tolerance range represents the range around the number of migration destination extrema contained in the migration destination extrema set.

7. The similarity calculation apparatus according to any one of claims 1 to 6, wherein, The similarity calculation device further includes a coordinate transformation unit, which calculates the coordinates obtained by projecting the extreme values ​​of each of the migration source data distributions (i.e., the migration source extreme values ​​before transformation) onto a space with a dimension higher than that of the migration source extreme values ​​before transformation, and uses these coordinates as the migration source extreme values. It also calculates the coordinates obtained by projecting the extreme values ​​of the migration destination data distributions (i.e., the migration destination extreme values ​​before transformation) onto a space with a dimension higher than that of the migration destination extreme values ​​before transformation, and uses these coordinates as the migration destination extreme values.

8. The similarity calculation apparatus according to any one of claims 1 to 6, wherein, The similarity calculation device further includes a data determination unit, which determines the migration source data candidates corresponding to the similarity that meets the determination conditions as migration source data.

9. The similarity calculation device according to claim 7, wherein, The similarity calculation device further includes a data determination unit, which determines the migration source data candidates corresponding to the similarity that meets the determination conditions as migration source data.

10. The similarity calculation apparatus according to any one of claims 1 to 6, wherein, The similarity calculation device further includes an extreme value calculation unit, which receives the migration source data distribution and the migration destination data distribution, uses the migration source data distribution to calculate the extreme value of the migration source data distribution, and uses the migration destination data distribution to calculate the extreme value of the migration destination data distribution.

11. The similarity calculation device according to claim 7, wherein, The similarity calculation device further includes an extreme value calculation unit, which receives the migration source data distribution and the migration destination data distribution, uses the migration source data distribution to calculate the extreme value of the migration source data distribution, and uses the migration destination data distribution to calculate the extreme value of the migration destination data distribution.

12. The similarity calculation device according to claim 8, wherein, The similarity calculation device further includes an extreme value calculation unit, which receives the migration source data distribution and the migration destination data distribution, uses the migration source data distribution to calculate the extreme value of the migration source data distribution, and uses the migration destination data distribution to calculate the extreme value of the migration destination data distribution.

13. The similarity calculation device according to claim 9, wherein, The similarity calculation device further includes an extreme value calculation unit, which receives the migration source data distribution and the migration destination data distribution, uses the migration source data distribution to calculate the extreme value of the migration source data distribution, and uses the migration destination data distribution to calculate the extreme value of the migration destination data distribution.

14. A similarity calculation apparatus, wherein the similarity calculation apparatus performs transfer learning using candidate transfer source data and candidate transfer destination data, wherein the candidate transfer source data and the candidate transfer destination data are training data for learning from Internet of Things (IoT) devices collected in different environments, and wherein the candidate transfer source data and the candidate transfer destination data are image data, wherein, The similarity calculation device has: The similarity calculation unit calculates the similarity between the source data distribution and the destination data distribution based on a set of migration source extreme values, including migration source extreme values ​​representing extreme values ​​in the source data distribution, and a set of destination extreme values, including migration destination extreme values ​​representing extreme values ​​in the destination data distribution. The source data distribution represents the distribution of feature quantities of the candidate source data, and the destination data distribution represents the distribution of feature quantities of the destination data. The extreme values ​​include maxima and minima. The coordinate transformation unit calculates the coordinates obtained by projecting the extreme values ​​of each of the migration source data distributions (i.e., the migration source extreme values ​​before transformation) onto a space with a dimension higher than that of the migration source extreme values ​​before transformation, and uses these coordinates as the migration source extreme values. It also calculates the coordinates obtained by projecting the extreme values ​​of the migration destination data distributions (i.e., the migration destination extreme values ​​before transformation) onto a space with a dimension higher than that of the migration destination extreme values ​​before transformation, and uses these coordinates as the migration destination extreme values.

15. The similarity calculation apparatus according to claim 14, wherein, The similarity calculation device further includes a data determination unit, which determines the migration source data candidates corresponding to the similarity that meets the determination conditions as migration source data.

16. The similarity calculation apparatus according to claim 14 or 15, wherein, The similarity calculation device further includes an extreme value calculation unit, which receives the migration source data distribution and the migration destination data distribution, uses the migration source data distribution to calculate the extreme value of the migration source data distribution, and uses the migration destination data distribution to calculate the extreme value of the migration destination data distribution.

17. A similarity calculation method, using candidate source data and destination data for transfer learning, wherein the candidate source data and destination data are training data collected from IoT devices in different environments for learning, and the candidate source data and destination data are image data, wherein... The similarity calculation unit calculates the similarity between the source data distribution and the destination data distribution based on a set of migration source extreme values ​​(including migration source extreme values ​​representing extreme values ​​in the source data distribution) and a set of destination extreme values ​​(including migration destination extreme values ​​representing extreme values ​​in the destination data distribution). The source data distribution represents the distribution of feature quantities of the candidate source data, and the destination data distribution represents the distribution of feature quantities of the destination data. The extreme values ​​include maxima and minima. There is at least one of each of the candidate migration source data, the migration source data distribution, and the migration source extreme value group. Each candidate migration source data corresponds one-to-one with each migration source data distribution, and each migration source data distribution corresponds one-to-one with each migration source extreme value group. The similarity calculation unit calculates the similarity between each of the migration source data distributions and the migration destination data distributions based on the migration source extreme value group and the migration destination extreme value group corresponding to each of the migration source data distributions. Each of the migration source extrema is associated with an identifier that identifies a migration source data candidate corresponding to the migration source data distribution. Based on the number of migration source extrema contained in the migration source extrema group, each of the migration source extrema belongs to at least one of the datasets prepared for each value represented by the number of extrema, wherein the migration source data distribution has extrema corresponding to each of the migration source extrema. The model generation unit generates a learning model corresponding to each of the datasets. This learning model is used to estimate the identifiers corresponding to the assigned coordinate sets. The model decision unit determines the learning model as the decision model based on the number of migration destination extrema contained in the migration destination extrema group. The similarity calculation unit uses the decision model and the extreme value set of the migration destination to calculate the similarity.

18. A similarity calculation method, using candidate source data and destination data for transfer learning, wherein the candidate source data and destination data are training data collected from IoT devices in different environments for learning, and the candidate source data and destination data are image data, wherein... The similarity calculation unit calculates the similarity between the source data distribution and the destination data distribution based on a set of migration source extreme values ​​(including migration source extreme values ​​representing extreme values ​​in the source data distribution) and a set of destination extreme values ​​(including migration destination extreme values ​​representing extreme values ​​in the destination data distribution). The source data distribution represents the distribution of feature quantities of the candidate source data, and the destination data distribution represents the distribution of feature quantities of the destination data. The extreme values ​​include maxima and minima. The coordinate transformation unit calculates the coordinates obtained by projecting the extreme values ​​of each of the migration source data distributions (i.e., the migration source extreme values ​​before transformation) onto a space with a dimension higher than that of the migration source extreme values ​​before transformation, and uses these coordinates as the migration source extreme values. It also calculates the coordinates obtained by projecting the extreme values ​​of the migration destination data distributions (i.e., the migration destination extreme values ​​before transformation) onto a space with a dimension higher than that of the migration destination extreme values ​​before transformation, and uses these coordinates as the migration destination extreme values.

19. A computer-readable recording medium containing a similarity calculation program that enables a computer to perform transfer learning using candidate source data and destination data, wherein the candidate source data and destination data are training data collected from Internet of Things (IoT) devices in different environments for learning, and wherein the candidate source data and destination data are image data, wherein... The similarity calculation program enables the computer to determine the similarity between the source data distribution and the destination data distribution based on a set of migration source extreme values ​​(including migration source extreme values ​​representing extreme values ​​in the source data distribution) and a set of destination extreme values ​​(including migration destination extreme values ​​representing extreme values ​​in the destination data distribution). The source data distribution represents the distribution of feature quantities of the candidate source data, and the destination data distribution represents the distribution of feature quantities of the destination data. The extreme values ​​include maxima and minima. There is at least one of each of the candidate migration source data, the migration source data distribution, and the migration source extreme value group. Each candidate migration source data corresponds one-to-one with each migration source data distribution, and each migration source data distribution corresponds one-to-one with each migration source extreme value group. The similarity calculation program enables the computer to determine the similarity between each of the source data distributions and the destination data distributions based on the source extreme value groups and destination extreme value groups corresponding to each of the source data distributions. Each of the migration source extrema is associated with an identifier that identifies a migration source data candidate corresponding to the migration source data distribution. Based on the number of migration source extrema contained in the migration source extrema group, each of the migration source extrema belongs to at least one of the datasets prepared for each value represented by the number of extrema, wherein the migration source data distribution has extrema corresponding to each of the migration source extrema. The similarity calculation procedure causes the computer to generate a learning model corresponding to each of the datasets. This learning model is used to estimate the identifiers corresponding to the assigned coordinate sets. The similarity calculation procedure causes the computer to determine the learning model as the deciding model based on the number of migration destination extrema contained in the migration destination extrema group. The similarity calculation program enables the computer to calculate the similarity using the decision model and the extreme value set of the migration destination.

20. A computer-readable recording medium containing a similarity calculation program that enables a computer to perform transfer learning using candidate source data and destination data, wherein the candidate source data and destination data are training data collected from Internet of Things (IoT) devices in different environments for learning, and wherein the candidate source data and destination data are image data, wherein... The similarity calculation program enables the computer to determine the similarity between the source data distribution and the destination data distribution based on a set of migration source extreme values ​​(including migration source extreme values ​​representing extreme values ​​in the source data distribution) and a set of destination extreme values ​​(including migration destination extreme values ​​representing extreme values ​​in the destination data distribution). The source data distribution represents the distribution of feature quantities of the candidate source data, and the destination data distribution represents the distribution of feature quantities of the destination data. The extreme values ​​include maxima and minima. The similarity calculation program enables the computer to calculate the coordinates obtained by projecting the extreme values ​​of each of the migration source data distributions (i.e., the migration source extreme values ​​before transformation) onto a space with a dimension higher than that of the migration source extreme values ​​before transformation, and to use these coordinates as the migration source extreme values. The program also calculates the coordinates obtained by projecting the extreme values ​​of the migration destination data distributions (i.e., the migration destination extreme values ​​before transformation) onto a space with a dimension higher than that of the migration destination extreme values ​​before transformation, and to use these coordinates as the migration destination extreme values.

Images