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

Abstract

To convert data into a common format between data groups expressed in different representation formats.SOLUTION: An information processing apparatus that converts data between at least two data groups expressed in different representation formats includes a data projection unit that projects a first data group and a second data group onto an N-dimensional hypersphere by using contrastive learning or the like, and data conversion unit that coverts the data such that distributions of the data groups projected onto the N-dimensional hypersphere match. The data projection unit projects the first data group and the second data group onto a narrower region on the N-dimensional hypersphere by initially setting a normalization coefficient of the contrastive learning to a large value, so that the data conversion unit can easily learn the match of the distributions between the data groups. The data projection unit enlarges a projection ranges of the first data group and the second data group on the N-dimensional hypersphere by setting the normalization coefficient to a smaller value gradually. The data conversion unit learns such that shapes of the data distributions match.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This invention relates to an information processing device, an information communication device, an information processing method, and an information processing program. [Background technology]

[0002] A technique has been proposed that generates a common feature vector shared by image data and text data, compares this common feature vector with brain activity feature vectors, and selectively supplies the brain activity feature vector to the channel that generates the vector with the highest correlation (see, for example, Patent Document 1). [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2018-205885 [Overview of the project] [Problems that the invention aims to solve]

[0004] In recent years, there has been a growing need in various fields to share or communicate data between multiple datasets. In such cases, the data representation formats may differ between datasets. For example, a dataset expressed in Japanese and one expressed in English, or datasets representing information perceived by different individuals' brains. To share or communicate information between datasets with such differing representation formats, it is necessary to convert the data into a common, format-independent representation format.

[0005] This invention was made in view of these circumstances, and its purpose is to convert data between data sets expressed in different representation formats into a common format that is independent of each representation format. [Means for solving the problem]

[0006] To solve the above problems, an information processing device in one aspect of the present invention is an information processing device that transforms data between at least two data sets expressed in different representation formats, comprising: a data projection unit that projects a first data set and a second data set onto an N-dimensional hypersphere using comparative learning, etc.; and a data transformation unit that transforms the data sets so that the distributions of the data sets projected onto the N-dimensional hypersphere match. The data projection unit initially sets the normalization coefficient of comparative learning to a large value, thereby projecting the first data set and the second data set onto narrow regions on the N-dimensional hypersphere, so that the data transformation unit can easily learn the match between the data sets. Subsequently, the data projection unit expands the projection range of the first data set and the second data set on the N-dimensional hypersphere by gradually setting the normalization coefficient to a small value. The data transformation unit learns so that the shapes of the data distributions match.

[0007] In one embodiment of the information processing device, comparative learning for the temperature parameter τ is performed so as to maximize the conditional probability P(i|v) given by equation (1). However, the feature representation vector on the N-dimensional hypersphere corresponding to the i-th data sample is f i Let v be the matrix consisting of the feature representation vectors of all data samples. T represents the transpose of the matrix.

number

number

number

[0008] In one embodiment of the information processing device, the data conversion unit may use Maximum Mean Discrepancy to determine the distance between the data distributions of a first data group and a second data group on an N-dimensional hypersphere.

[0009] In one embodiment of the information processing device, the data conversion unit may use a generative adversarial network to learn that the shapes of the data distributions of the first data group and the second data group on an N-dimensional hypersphere match.

[0010] An information processing device in one embodiment may further include a first feature representation vector generation unit that generates a first feature representation vector from a first data group, and a second feature representation vector generation unit that generates a second feature representation vector from a second data group. In this case, the data projection unit projects the first feature representation vector and the second feature representation vector onto an N-dimensional hypersphere. The data transformation unit transforms the first feature representation vector and the second feature representation vector projected onto the N-dimensional hypersphere into first and second common variable representation vectors, respectively, by first and second transformations. The first and second common variable representation vectors are converted back into the first and second feature representation vectors, respectively, by inverse transformations of the first and second transformations.

[0011] Another aspect of the present invention is an information communication device. This device is an information communication device for communicating data between at least two data sets expressed in different representation formats, and comprises: a data projection unit that projects a first data set and a second data set onto an N-dimensional hypersphere using comparative learning, etc.; a data transformation unit that transforms the data sets projected onto the N-dimensional hypersphere so that their distributions match; and a data communication unit that transmits and receives the transformed data. The data projection unit initially sets the normalization coefficient of comparative learning to a large value, thereby projecting the first data set and the second data set onto narrow regions on the N-dimensional hypersphere, respectively, so that the data transformation unit can easily learn the overlap of the data sets. Subsequently, the data projection unit expands the projection range of the first data set and the second data set on the N-dimensional hypersphere by gradually setting the normalization coefficient to a small value. The data transformation unit learns so that the shapes of the data distributions match.

[0012] A further aspect of the present invention is an information processing method. This method is an information processing method for transforming data between at least two data sets represented in different representation formats, and includes a data projection step of projecting a first data set and a second data set onto an N-dimensional hypersphere, for example, by using symmetric learning, and a data transformation step of transforming the data sets so that the distributions of the projected data sets on the N-dimensional hypersphere coincide. The data projection step projects the first data set and the second data set onto narrow regions on the N-dimensional hypersphere by initially setting the normalization coefficient of symmetric learning to a large value. Subsequently, the data projection step expands the projection range of the first data set and the second data set on the N-dimensional hypersphere by gradually setting the normalization coefficient to a small value. The data transformation step learns so that the shapes of the data distributions coincide.

[0013] Another aspect of the present invention is an information processing program. This program is an information processing program for converting data between at least two data groups expressed in different expression forms, and includes a data projection step of projecting a first data group and a second data group onto an N-dimensional hypersphere surface using contrast learning or the like, and a data conversion step of converting the data groups projected onto the N-dimensional hypersphere surface so that their distributions match, and causing a computer to execute them. In the data projection step, by first setting the normalization coefficient of contrast learning to a large value, the first data group and the second data group are projected onto narrow regions on the N-dimensional hypersphere surface respectively. Then, in the data projection step, by gradually setting the normalization coefficient to a small value, the projection ranges of the first data group and the second data group on the N-dimensional hypersphere surface are expanded. The data conversion step learns so that the shapes of the data distributions of the data groups match.

[0014] In addition, any combination of the above components, and those obtained by converting the expression of the present invention among devices, methods, systems, recording media, computer programs, etc., are also effective as aspects of the present invention.

Advantages of the Invention

[0015] According to the present invention, data can be converted into a common form independent of each expression form between data groups expressed in different expression forms.

[0016] Also according to the present invention, one data can be converted into the other expression form between data groups expressed in different expression forms. The state at this time is shown in FIG. 7.

Brief Description of the Drawings

[0017] [Figure 1] It is a functional block diagram of an information processing apparatus according to the first embodiment. [Figure 2] It is a diagram showing the data distribution immediately after the data is projected onto the N-dimensional hypersphere surface (that is, in the initial state). [Figure 3]This figure shows how the first and second data sets are projected onto narrow regions on an N-dimensional hypersphere by initially setting the normalization coefficient Z to a large value (i.e., Z=104). [Figure 4] This figure shows how gradually setting the normalization coefficient to a smaller value (i.e., Z=10²) expands the projection range of the first and second data sets on the N-dimensional hypersphere, and how the data transformation unit learns to match the shape of the data distributions (the second data set is displayed after being transformed into the representation format of the first data set). [Figure 5] This is a functional block diagram of the information processing device according to the second embodiment. [Figure 6] This figure schematically illustrates the operation of the information processing device according to the second embodiment. [Figure 7] This diagram schematically illustrates the process of converting data from one representation format to another when the data is represented in different formats. [Figure 8] This is a functional block diagram of an information and communication device according to the third embodiment. [Figure 9] This flowchart shows the processing procedures for the information processing method according to the fourth embodiment and the information processing program according to the fifth embodiment. [Modes for carrying out the invention]

[0018] The present invention will be described below with reference to the drawings, based on preferred embodiments. In embodiments and modifications, the same or equivalent components and members will be denoted by the same reference numerals, and redundant explanations will be omitted as appropriate. In addition, the dimensions of the members in each drawing will be enlarged or reduced as appropriate to facilitate understanding. Furthermore, some members that are not important for explaining the embodiments will be omitted in each drawing. In addition, terms including ordinal numbers such as "first," "second," etc., are used to describe various components, but these terms are used only to distinguish one component from others, and the components are not limited by these terms.

[0019] [First Embodiment] Figure 1 is a functional block diagram of the information processing device 1 according to the first embodiment. The information processing device 1 comprises a data projection unit 10 and a data conversion unit 20.

[0020] The information processing device 1 is connected to a first data group and a second data group. The first and second data groups are sets of data consisting of data expressed in different representation formats. Examples include a data group of an object expressed in Japanese and a data group of an object expressed in English, or a data group of information about an object perceived by the brains of different people. Because the representation formats of the first and second data groups are different, it is not possible to directly share or communicate data between these data groups.

[0021] The first data set and the second data set are input to the data projection unit 10 of the information processing device 1. The data projection unit 10 projects the first data set and the second data set onto an N-dimensional hypersphere, using methods such as comparative learning.

[0022] As mentioned above, the first and second data sets have different data representation formats. Therefore, when these data are projected onto an N-dimensional hypersphere, they appear to have completely different data distributions. However, these data may share a common conceptual structure. For example, if the first data set is a collection of data expressed in Japanese and the second data set is a collection of data expressed in English, concepts such as "cat," "lion," and "car" will be projected to different positions on the N-dimensional hypersphere. However, in both data sets, similar concepts such as "cat" and "lion" are represented as vectors in close proximity. On the other hand, dissimilar concepts such as "cat" and "car" are represented as vectors in distant positions. Thus, even if data sets have different representation formats, if the relationship structure between concepts has common properties, it is possible to rewrite the data using a common representation format.

[0023] The data projection unit 10 projects the data onto an N-dimensional hypersphere using methods such as symmetric learning. Symmetric learning is a machine learning technique that learns to project similar data closer together and dissimilar data further away.

[0024] In contrastive learning, the distance between similar data points on an N-dimensional hypersphere is minimized while the distance between different data points on an N-dimensional hypersphere is maximized. If we want to learn the separation of individual data samples, the learning process is performed to maximize the conditional probability described by equation (1).

number

[0025] The data projection unit 10 initially projects the first data group and the second data group onto a single point or a narrow range on an N-dimensional hypersphere by setting the normalization coefficient for comparative learning to a large value, and then learns to broaden the data distribution range of the first data group and the second data group on the N-dimensional hypersphere by gradually setting the normalization coefficient to a smaller value.

[0026] The data transformation unit 20 transforms the first and second data sets so that their distributions, projected onto an N-dimensional hypersphere, coincide. When both data sets are projected onto a single point or a narrow range, it is relatively easy to transform them so that they overlap. Subsequently, by gradually setting the normalization coefficient of the data projection unit 10 to a smaller value, the data transformation unit 20 learns to make the two data distributions coincide as the data distribution ranges of the first and second data sets on the N-dimensional hypersphere are expanded.

[0027] Figures 2 to 4 illustrate the operation of the data projection unit 10, which projects two data sets onto an N-dimensional hypersphere using comparative learning, and the data transformation unit 20, which transforms the two data sets so that their distributions match. In these figures, the transformed data is generated by matching the second data set to the data format of the first data set. In Figures 2 to 4, the left figure shows the data distribution of the first data set, the right figure shows the data distribution of the second data set, and the center figure shows the result of matching the second data set to the representation format of the first data set (data transformation).

[0028] Figure 2 shows the data distribution immediately after the data projection unit 10 projects the data onto an N-dimensional hypersphere (i.e., the initial state). At first glance, the distributions of the first data group (left figure) and the second data group (right figure) appear to be quite different. The center figure shows the data distribution (the transformed data representation format in the initial state) obtained when the second data group is transformed onto the representation format of the first data group.

[0029] Figure 3 shows that the data projection unit 10 initially sets the normalization coefficient Z to a large value (i.e., Z=10) 4By setting it to ), the first and second data sets are represented as data sets distributed in narrow regions on an N-dimensional hypersphere. The first data set (left figure) and the second data set (right figure) are mapped to highly condensed regions on the N-dimensional hypersphere (the first data set is represented as thin lines, and the second data set as almost points). This makes it very easy for the data transformation unit 20 to learn the correspondence between the two sets of data. The center figure shows how the second data set has been transformed into a representation of the first data set in a very small region (very short and thin lines).

[0030] Figure 4 shows that the data projection unit 10 gradually reduces the normalization coefficient to a smaller value (i.e., Z=10). 2 By setting it to ), the distributions of the first and second data sets on the N-dimensional hypersphere are projected in a way that expands them. By relaxing the normalization coefficient from a large value to a small value, the first and second data sets are extended to a larger area on the N-dimensional hypersphere. The data transformation unit 20 learns so that the shapes of the two distributions match. The central figure shows how the second data set has been transformed into a representation of the first data set that extends to a wider area.

[0031] As described above, the data projection unit 10 can represent the first and second data groups as data groups in narrow regions on an N-dimensional hypersphere by initially setting the normalization coefficient for comparative learning to a large value. This allows the data transformation unit 20 to easily learn the agreement between the two representation formats. Subsequently, the data projection unit 10 expands the projection range of the first and second data groups on the N-dimensional hypersphere by gradually setting the normalization coefficient to a smaller value. By learning that the shapes of the two data distributions match, the data transformation unit 20 can match the representation formats of the first and second data groups with high accuracy.

[0032] Another method for matching data sets with different representation formats involves performing comparative learning on a multidimensional hypersphere, mapping highly similar elements close together and less similar elements far apart. This is followed by extracting corresponding elements called anchor points, and then achieving matching of these anchor points and the overall matching of the two data distributions using a machine learning technique. However, this method has a problem: if elements that are not corresponding are extracted as anchor points, the accuracy of conversion between data representations and the accuracy of extracting common information representations are significantly reduced. Furthermore, the anchor point extraction process requires information processing independent of the machine learning technique, including comparative learning, making it inefficient.

[0033] In contrast, this embodiment enables data conversion between different data representation formats and extraction of common representations without requiring the extraction of anchor points. Furthermore, since the objective can be achieved simply by manipulating normalization coefficients within the same machine learning process, computational efficiency is dramatically improved. Moreover, because it does not depend on the accuracy of anchor point extraction, it is possible to obtain effects such as improved accuracy in data conversion and extraction of common representations.

[0034] In one embodiment of the present invention, the condition probability P(i|v) for which control learning should be maximized at the temperature parameter τ is

number

[0035] According to this embodiment, the conditional probability that should be maximized by controlled learning can be specifically defined.

[0036] The data conversion unit 20 may use Maximum Mean Discrepancy (MMD) to determine the distance between the data distributions of the first data group and the second data group on the N-dimensional hypersphere.

[0037] According to this embodiment, the distance between data distributions can be determined more accurately.

[0038] The data transformation unit may use a generative adversarial network (GAN) to learn how to make the shapes of the data distributions of the first and second data sets on an N-dimensional hypersphere match.

[0039] According to this embodiment, learning of matching data distribution shapes can be performed more efficiently. Other methods may also be used to learn matching data distribution shapes.

[0040] [Second Embodiment] Figure 5 is a functional block diagram of the information processing device 2 according to the second embodiment. The information processing device 2 comprises a data projection unit 10, a data conversion unit 20, a first feature representation vector generation unit 30, and a second feature representation vector generation unit 40. In other words, the information processing device 2 comprises the first feature representation vector generation unit 30 and the second feature representation vector generation unit 40 in addition to the configuration of the information processing device 1 in Figure 1. The other configurations of the information processing device 2 are the same as those of the information processing device 1.

[0041] The first feature representation vector generation unit 30 generates a first feature representation vector from the first data set. The second feature representation vector generation unit 40 generates a second feature representation vector from the second data set.

[0042] The data projection unit 10 projects the first feature representation vector and the second feature representation vector onto an N-dimensional hypersphere, respectively.

[0043] The data transformation unit 20 transforms the first and second feature representation vectors projected onto the N-dimensional hypersphere into first and second common variable representation vectors using the first transformation F and the second transformation G, respectively. The first and second common variable representation vectors are then converted back into the first and second feature representation vectors, respectively, using the inverse transformations F' and G' of the first and second transformations F and G'.

[0044] Figure 6 schematically shows the operation of the information processing device 2.

[0045] The first feature representation vector generation unit 30 generates a first feature representation vector corresponding to the representation format included in the first data group. For example, when the first data group consists of n data, the first feature representation vector generation unit 30 generates n first feature representation vectors Sa = [Sa_1, Sa_2, …, Sa_n] from these n data. Here, n is a natural number.

[0046] Similarly, the second feature representation vector generation unit 40 generates a second feature representation vector corresponding to the representation format included in the second data group. For example, when the second data group consists of m data, the second feature representation vector generation unit 40 generates m second feature representation vectors Sb = [Sb_1, Sb_2, …, Sb_m] from these m data. Here, m is a natural number.

[0047] The data projection unit 10 projects the first feature representation vector Sa and the second feature representation vector Sb onto an N-dimensional hypersphere, respectively.

[0048] The data conversion unit 20 converts the first feature representation vector Sa and the second feature representation vector Sb projected onto the N-dimensional hypersphere into first and second common variable representation vectors Sa’ = F · Sa = [Sa_1’, Sa_2’, …, Sa_n’] Sb’ = G · Sb = [Sb_1’, Sb_2’, …, Sb_m’] by the first conversion F and the second conversion G, respectively. The first and second common variable representation vectors Sa’ and Sb’ return to the first feature representation vector Sa and the second feature representation vector Sb by the inverse conversions F’ and G’ of the first conversion F and the second conversion G, respectively. That is, Sa = F -1 · G · G -1 · F · Sa Sb = G -1 · F · F -1 · G · Sb This is the result.

[0049] According to this embodiment, data conversion between different data representation formats can be performed more accurately.

[0050] [Third Embodiment] Figure 8 is a functional block diagram of the information communication device 3 according to the third embodiment. The information communication device 3 comprises a data projection unit 10, a data conversion unit 20, and a data communication unit 50. In other words, the information communication device 3 further comprises the data communication unit 50 in addition to the configuration of the information processing device 1 shown in Figure 1. The other configurations of the information communication device 3 are the same as those of the information processing device 1.

[0051] The data communication unit 50 takes the first data group as input, passes it through the data projection unit 10, and transmits the data whose representation format has been converted by the data conversion unit 20 to the information source of the second data group. It then takes the second data group as input, passes it through the data projection unit 10, and transmits the data whose representation format has been converted by the data conversion unit 20 to the information source of the first data group.

[0052] According to this embodiment, data can be sent and received between systems that handle data sets with different representation formats.

[0053] [Fourth Embodiment] Figure 9 is a flowchart showing the processing procedure of an information processing method according to the fourth embodiment. This method includes a data projection step S10 and a data conversion step S20.

[0054] The data projection step S10 projects the first and second data sets, which are represented in different representation formats, onto an N-dimensional hypersphere using methods such as symmetric learning.

[0055] Next, in the data transformation step S20, the data is transformed so that the distributions of the data sets projected onto the N-dimensional hypersphere match.

[0056] The data projection step S10 projects the first and second data sets onto narrow regions on an N-dimensional hypersphere by initially setting the normalization coefficient for symmetric learning to a large value. Subsequently, the data projection step S10 expands the projection range of the first and second data sets on the N-dimensional hypersphere by gradually setting the normalization coefficient to a smaller value. The data transformation step S20 learns so that the shapes of the data distributions match.

[0057] According to this embodiment, data conversion between different data representation formats can be performed using a computer.

[0058] [Fifth Embodiment] The fifth embodiment is an information processing program. This program causes a computer to execute each step of the processing flow shown in Figure 9.

[0059] The data projection step S10 projects the first and second data sets, which are represented in different representation formats, onto an N-dimensional hypersphere using methods such as symmetric learning.

[0060] Next, in the data transformation step S20, the data is transformed so that the distributions of the data sets projected onto the N-dimensional hypersphere match.

[0061] The data projection step S10 projects the first and second data sets onto narrow regions on an N-dimensional hypersphere by initially setting the normalization coefficient for symmetric learning to a large value. Subsequently, the data projection step S10 expands the projection range of the first and second data sets on the N-dimensional hypersphere by gradually setting the normalization coefficient to a smaller value. The data transformation step S20 learns so that the shapes of the data distributions match.

[0062] According to this embodiment, a program that performs data conversion between different data representation formats can be implemented as computer software.

[0063] The present invention has been described above based on embodiments. These embodiments are illustrative, and it will be understood by those skilled in the art that various modifications are possible in combinations of their components and processing processes, and that such modifications also fall within the scope of the present invention.

[0064] In understanding the technical concept abstracted from the embodiments and modifications, that technical concept should not be interpreted restrictively to the content of the embodiments and modifications. The embodiments and modifications described above are merely examples, and many design changes, such as changes, additions, and deletions of components, are possible. In the embodiments, the content in which such design changes are possible is emphasized with the notation "embodiment." However, design changes are also permitted in content without such notation. [Explanation of Symbols]

[0065] 1. Information processing device, 2. Information processing device, 3. Information and communication equipment, 10. Data projection unit, 20. Data conversion unit, 30. First feature representation vector generation unit, 40. Second feature representation vector generation unit, 50. Data Communications Department, S10 · Data projection step, S20...Data conversion step, Sa··First feature representation vector, Sb··Second feature representation vector, Sa'··First common variable representation vector, Sb'...Second common variable representation vector.

Claims

1. An information processing device for converting data between at least two data sets expressed in different representation formats, A data projection unit that projects the first and second data sets onto an N-dimensional hypersphere using comparative learning, A data transformation unit that transforms the data sets projected onto the N-dimensional hypersphere so that their distributions match, Equipped with, The data projection unit is, By initially setting the normalization coefficient for comparative learning to a large value, the first and second data sets are projected onto narrow regions on an N-dimensional hypersphere, allowing the data transformation unit to easily learn the agreement of the distributions between the data sets. Subsequently, by gradually setting the normalization coefficient to a smaller value, the projection range of the first and second data sets on the N-dimensional hypersphere is expanded. The data conversion unit is characterized by learning to match the shape of the data distribution.

2. When performing controlled learning on the temperature parameter τ, the conditional probability P(i|v) that should be maximized is [Math 3] The information processing apparatus according to claim 1, characterized in that it manipulates the normalization coefficient Z.

3. The information processing apparatus according to claim 1 or 2, characterized in that the data conversion unit determines the distance between the data distributions of the first data group and the second data group on the N-dimensional hypersphere using Maximum Mean Discrepancy.

4. The information processing apparatus according to claim 1 or 2, characterized in that the data conversion unit uses a generative adversarial network to learn that the shapes of the data distributions of the first data group and the second data group on the N-dimensional hypersphere match.

5. A first feature representation vector generation unit generates a first feature representation vector from the first data set, A second feature representation vector generation unit generates a second feature representation vector from the second data set, Furthermore, The data projection unit projects the first feature representation vector and the second feature representation vector onto the N-dimensional hypersphere using comparative learning, The data conversion unit converts the first feature representation vector and the second feature representation vector projected onto the N-dimensional hypersphere into a first common variable representation vector and a second common variable representation vector, respectively, by the first and second transformations. The information processing apparatus according to claim 1 or 2, characterized in that the first common variable representation vector and the second common variable representation vector are returned to the first feature representation vector and the second feature representation vector, respectively, by the inverse transformations of the first and second transformations.

6. An information communication device that communicates data between at least two data sets expressed in different representation formats, A data projection unit that projects the first and second data sets onto an N-dimensional hypersphere using comparative learning, A data transformation unit that transforms the data sets projected onto the N-dimensional hypersphere so that their distributions match, A data communication unit that transmits and receives converted data, Equipped with, The data projection unit is, By initially setting the normalization coefficient for comparative learning to a large value, the first and second data sets are projected onto narrow regions on an N-dimensional hypersphere, allowing the data transformation unit to easily learn the overlap of the data sets. Subsequently, by gradually setting the normalization coefficient to a smaller value, the projection range of the first and second data sets on the N-dimensional hypersphere is expanded. The data conversion unit is characterized by learning to match the shape of the data distribution.

7. An information processing method for converting data between at least two sets of data expressed in different representation formats, A data projection step in which the first and second data sets are projected onto an N-dimensional hypersphere using symmetric learning, A data transformation step is to transform the data so that the distributions of the data sets projected onto the N-dimensional hypersphere coincide, Includes, The aforementioned data projection step is, By initially setting the normalization coefficient for comparative learning to a large value, the first and second data sets are projected onto narrow regions on an N-dimensional hypersphere, allowing the data transformation unit to easily learn the agreement of the distributions between the data sets. Subsequently, by gradually setting the normalization coefficient to a smaller value, the projection range of the first and second data sets on the N-dimensional hypersphere is expanded. The data transformation step is characterized by learning to match the shape of the data distribution.

8. An information processing program that converts data between at least two sets of data expressed in different representation formats, A data projection step in which the first and second data sets are projected onto an N-dimensional hypersphere using symmetric learning, A data transformation step is to transform the data so that the distributions of the data sets projected onto the N-dimensional hypersphere coincide, Have the computer run it, The aforementioned data projection step is, By initially setting the normalization coefficient for comparative learning to a large value, the first and second data sets are projected onto narrow regions on an N-dimensional hypersphere, allowing the data transformation unit to easily learn the agreement of the distributions between the data sets. Subsequently, by gradually setting the normalization coefficient to a smaller value, the projection range of the first and second data sets on the N-dimensional hypersphere is expanded. The data transformation step is characterized by learning to match the shape of the data distribution.

Images