Self-supervised learning device for artificial intelligence algorithm, and method therefor

Abstract

The present invention relates to a device for training an artificial intelligence algorithm, and a method therefor and, more specifically, to a device for training an artificial intelligence algorithm through a preset self-supervised learning method, and a method therefor. The self-supervised learning device for an artificial intelligence algorithm according to one embodiment of the present invention comprises: a memory in which original data and an artificial intelligence algorithm are stored; and a processor, wherein the memory can store program instructions, that can be executed by the processor, of: inputting learning data corresponding to the original data into a first feature extractor and a second feature extractor of the artificial intelligence algorithm; generating 1-1 feature information and 1-2 feature information by using results of the first feature extractor; generating 2-1 feature information and 2-2 feature information by using results of the second feature extractor; and training the second feature extractor by using the 1-1 feature information, the 1-2 feature information, the 2-1 feature information, and the 2-2 feature information.

Description

Self-supervised learning device and method for artificial intelligence algorithms

[0001] The present invention relates to a device and method for training an artificial intelligence algorithm, and more specifically, to a device and method for training an artificial intelligence algorithm through a preset self-supervised learning method.

[0002] Supervised learning is the most widely used method for training AI algorithms. Supervised learning refers to a method in which AI algorithms learn using training data with labels (correct answers). However, if the data with correct answers is missing or insufficient, AI algorithms either cannot learn or their learning effectiveness is limited.

[0003] To address this, self-supervised learning has been introduced to AI algorithm learning. Self-supervised learning refers to an AI learning method that extracts feature vectors from input data itself, without the need for separate labels. For example, in contrastive learning, feature vectors are calculated from three types of data: (1) original data, (2) transformed data, and (3) data different from the original data. The AI ​​algorithm can be trained to ensure that (1) and (2) have similar values, while (3) has a different value. However, self-supervised learning can have the problem of forcing (2) and (3) to have different feature vectors even when they have similar information.

[0004] To address this, a similarity distillation-based approach has recently been introduced to AI algorithm learning. Similarity distillation utilizes not only (2) and (3) but also a pre-prepared data queue. The data queue can consist of tens or thousands of data items (e.g., image data). Furthermore, while the data queue may differ from the original data, it may also contain similar data. Similarity distillation trains a student model so that the calculated feature vectors derived from a pre-trained "teacher model" (teacher encoder) are identical to the calculated feature vectors derived from the target "student model" (student encoder). The name "similarity distillation" derives from the similarity between the teacher model and the student model, which transfers "similarity-related knowledge" to the student model, reminiscent of the process of distillation in a chemistry lab. This similarity distillation approach considers one-dimensional relationships. However, research in the field of graphs has proven that accuracy increases as higher-dimensional relationships are considered.

[0005] Therefore, an artificial intelligence algorithm learning method that can improve accuracy by taking into account higher-dimensional relationships is required.

[0006] In order to solve the above-described problems, the present invention provides a device and method capable of training an artificial intelligence algorithm through high-dimensional relationships.

[0007] According to one embodiment of the present invention, a self-supervised learning device is disclosed, comprising: a memory in which original data and an artificial intelligence algorithm are stored; and a processor; wherein the memory stores program commands that are executable by the processor, inputting learning data corresponding to the original data into a first feature extractor and a second feature extractor of the artificial intelligence algorithm, generate 1-1 feature information and 1-2 feature information using a result of the first feature extractor, generate 2-1 feature information and 2-2 feature information using a result of the second feature extractor, and train the second feature extractor using the 1-1 feature information, the 1-2 feature information, the 2-1 feature information, and the 2-2 feature information.

[0008] According to an embodiment, the memory may store program instructions executable by the processor to generate first learning data and second learning data using the original data, input the first learning data to the first feature extractor, and input the second learning data to the second feature extractor.

[0009] According to an embodiment, the memory may store program instructions that generate at least one of the first learning data and the second learning data through one or more of color change, rotation, and partial selection of the original data, executable by the processor.

[0010] In an embodiment, the memory further includes n data queues, and stores program instructions executable by the processor to generate the 1-1 feature information, the 1-2 feature information, the 2-1 feature information, and the 2-2 feature information using a first feature vector, a second feature vector, and a feature vector of each of the n data queues, wherein the first feature vector is an output value of the first feature extractor, the second feature vector is an output value of the second feature extractor, and the feature vector of each of the n data queues is a preset value or information previously generated through the first feature extractor, and n may be a natural number greater than or equal to 2.

[0011] According to an embodiment, the memory may store program instructions that are executable by the processor and produce the first-first characteristic information by Equation 1 through Equation 4.

[0012] [Mathematical Formula 1]

[0013]

[0014] [Mathematical Formula 2]

[0015]

[0016] [Equation 3]

[0017]

[0018] [Equation 4]

[0019]

[0020] However, the above Z is a first feature matrix of (n+2) dimensions, and is the transpose matrix of the first feature matrix, A is an adjacency matrix of dimension (n+2)*(n+2), s is a sigmoid function, and is a connectivity-based matrix of (n+2) dimensions corresponding to the first feature vector, and and above are the weight values ​​of the first MLP stored in advance, and the above are the values ​​corresponding to the first feature vector in the above adjacency matrix, and is the above 1-1 feature information, and is a (n+2)-dimensional matrix corresponding to the first feature vector, and and above are the weight values ​​of the second MLP that were previously stored.

[0021] According to an embodiment, the memory stores program instructions that are executable by the processor and produce the second-first characteristic information by the mathematical expression 1 and the mathematical expression 4, is a connectivity-based matrix of (n+2) dimensions corresponding to the second feature vector, and and above are the weight values ​​of the first MLP stored in advance, and the above are the values ​​corresponding to the second feature vector in the above adjacency matrix, and is the above 2-1 feature information, and is a (n+2)-dimensional matrix corresponding to the second feature vector, and and above may be the weight values ​​of the second MLP stored in advance.

[0022] According to an embodiment, the memory may store program instructions executable by the processor for producing the first feature matrix as the first-second feature information and the second-second feature information.

[0023] According to an embodiment, the memory may store program instructions executable by the processor, which generate first comparison information using the first-1 feature information and the first-2 feature information, generate second comparison information using the second-1 feature information and the second-2 feature information, and train the second feature extractor by comparing the first comparison information and the second comparison information.

[0024] According to an embodiment, the first comparison information may be a sum of the 1-1 feature information and the 1-2 feature information, and the second comparison information may be a sum of the 2-1 feature information and the 2-2 feature information.

[0025] According to an embodiment, the memory may store program instructions executable by the processor that train the second feature extractor so that the similarity distribution of the second comparison information follows the similarity distribution of the first comparison information.

[0026] The self-supervised learning device and method of the artificial intelligence algorithm according to the present invention can train the artificial intelligence algorithm by considering not only the direct connectivity (first-order connectivity) between the original data and the data queue, but also the connectivity (higher-order connectivity) between the data queues, thereby improving accuracy.

[0027] In order to more fully understand the drawings cited in the detailed description of the present invention, a brief description of each drawing is provided.

[0028] FIG. 1 is a schematic diagram illustrating a self-directed learning device according to one embodiment of the present invention.

[0029] Figure 2 is a flowchart for explaining the operation of a self-directed learning device according to one embodiment of the present invention.

[0030] FIG. 3 is a diagram for explaining original data, first learning data, and second learning data according to one embodiment of the present invention.

[0031] FIG. 4 is a diagram for explaining the operation of generating first characteristic information and second characteristic information according to one embodiment of the present invention.

[0032] FIG. 5 is a diagram for explaining the effect of an artificial intelligence algorithm learned according to one embodiment of the present invention.

[0033] The technical concept of the present disclosure is susceptible to various modifications and various embodiments. Specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the technical concept of the present disclosure to specific embodiments, and it should be understood that all modifications, equivalents, and alternatives fall within the scope of the technical concept of the present disclosure.

[0034] When explaining the technical concepts of the present disclosure, detailed descriptions of related known technologies will be omitted if they are deemed to unnecessarily obscure the gist of the technical concepts of the present disclosure. Furthermore, numbers (e.g., "first," "second," etc.) used in the description of the present disclosure are merely identifiers used to distinguish one component from another.

[0035] Additionally, when a component is referred to herein as being "connected" or "connected" to another component, it should be understood that the component may be directly connected or connected to the other component, but may also be connected or connected via another component in between, unless otherwise specifically stated.

[0036] In addition, terms such as “~part,” “~device,” and “~sub-unit” described herein mean a unit that processes at least one function or operation, which may be implemented by hardware such as a processor, a microprocessor, a microcontroller, a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processor unit (APU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), software, or a combination of hardware and software.

[0037] It should be noted that the division of components herein is merely a distinction based on the primary function each component is responsible for. In other words, two or more components described below may be combined into a single component, or a single component may be further subdivided into two or more components with more detailed functions. Furthermore, each component described below may, in addition to its own primary function, additionally perform some or all of the functions performed by other components. Furthermore, it should be noted that some of the primary functions of each component may be exclusively performed by other components.

[0038] Hereinafter, various embodiments according to the technical idea of ​​the present disclosure will be described in detail.

[0039]

[0040] FIG. 1 is a schematic diagram illustrating a self-directed learning device according to one embodiment of the present invention.

[0041] Referring to FIG. 1, a self-directed learning device (100) according to one embodiment of the present invention may include a memory (MEMORY, 110) and a processor (PROCESSOR, 120).

[0042] The memory (110) is a configuration in which various information and program commands for the operation of the text recognition device (100) are stored, and may be a storage device such as a hard disk, an SSD (Solid State Drive), etc. In particular, the memory (110) may store an artificial intelligence algorithm. In addition, the memory (110) may store original data for self-supervising the artificial intelligence algorithm. Here, the original data may be multimedia data such as images, videos, etc. In addition, the memory (110) may store program commands for executing 'artificial intelligence algorithm self-supervised learning' that can be executed by the processor (120).

[0043] The processor (120) is a central processing unit (CPU) capable of controlling the overall operation of the self-supervised learning device (100), and in particular, processes original data using artificial intelligence algorithms and other program commands stored in the memory (110), and can perform self-supervised learning on the artificial intelligence algorithm using the processed original data. Hereinafter, the self-supervised learning operation under the control of the processor (120) will be described.

[0044]

[0045] Figure 2 is a flowchart for explaining the operation of a self-directed learning device according to one embodiment of the present invention.

[0046] Referring to FIG. 2, the processor (120) can perform self-supervised learning (200) on an artificial intelligence algorithm stored in the memory (110). The artificial intelligence algorithm according to one embodiment of the present invention may be self-supervised learning by a similarity distillation method. For example, through the artificial intelligence algorithm according to one embodiment of the present invention, the feature vectors of the first learning data and n pre-stored data queues (where n is a natural number greater than or equal to 2) can all be calculated in the 'teacher model (teacher encoder)', and the feature vector of the second learning data can be calculated in the 'student model (student encoder)', which is a learning target. In addition, the similarity distributions of the teacher model and the student model can be calculated through the feature vectors calculated in the teacher model and the feature vectors calculated in the student model, and the student model can be trained so that the calculated similarities become the same. Hereinafter, each operation of FIG. 2 will be described in more detail. Each operation described below may be performed by executing program commands under the control of the processor (120), but for the convenience of understanding and explanation, it is described as being performed by the processor (120).

[0047]

[0048] First, the processor (120) can process the original data (ORIGINAL DATA, 210) according to a preset method to generate first learning data (1st LEARNING DATA) and second learning data (2nd LEARNING DATA). The first learning data (220-1) may be learning data input to the teacher model (230-1), and the second learning data (220-2) may be learning data input to the student model (230-2).

[0049] FIG. 3 is a diagram for explaining original data, first learning data, and second learning data according to one embodiment of the present invention. Referring to FIG. 3, an example is provided in which the processor (120) selects a portion of the original data (300) to generate first learning data (310) and second learning data (320). That is, the processor (120) may extract only the face portion of the cat, which is the subject of the original data (300), to a preset size to generate the first learning data (310). In addition, the processor (120) may extract only the ear portion of the cat, which is the subject of the original data (330), to a preset size to generate the second learning data (320). At this time, the method for generating the first learning data and / or the second learning data may vary.

[0050] For example, as illustrated in FIG. 3, areas corresponding to multiple learning data ('areas 1' to 'areas 8' in FIG. 3) may be preset in the original data (300), and the processor (120) may select some of the multiple areas and utilize them to generate learning data.

[0051] For another example, the processor (120) may apply one or more preset image conversion techniques to the original data (300) to generate different images as learning data (310, 320). Examples of the image conversion techniques may include, but are not limited to, the following.

[0052] (1) Technology for randomly cutting out a portion of the original data (300) (crop)

[0053] (2) Technology for rotating original data (300) at different angles (rotation)

[0054] (3) Technology for smoothing the original data (300) (smoothing or blurring)

[0055] (4) Technology for mixing random noise signals into original data (300)

[0056]

[0057] Thereafter, the processor (120) can use the generated learning data to perform self-supervised learning on an artificial intelligence algorithm. The self-supervised learning operation (200) of the artificial intelligence algorithm is illustrated within the dotted box in FIG. 2.

[0058] The self-supervised learning operation (200) of the artificial intelligence algorithm may include logical configurations such as a first feature extractor (1st FEATURE EXTRACTOR, 230-1), a second feature extractor (2nd FEATURE EXTRACTOR, 230-2), a first multi-layer perceptron (1st MLP, 1st Multi-Layer Perceptron, 240-1), a second multi-layer perceptron (2nd MLP, 2nd Multi-Layer Perceptron, 240-2), a graph generator (GRAPH GENERATOR, 250), a structural feature generator (STRUCTURE BASED EMBEDDING, 260-1), a node feature generator (NODE FEATURE BASED EMBEDDING, 260-2), and a Kullback-Leibler divergence loss (KL DIVERGENCE LOSS, 270). Hereinafter, individual logical configurations of a self-supervised learning operation (200) of an artificial intelligence algorithm according to one embodiment of the present invention will be described in detail. Each logical configuration (230-1 to 270) to be described below may be program instructions set to perform a preset function according to the execution of the processor (120). However, for the convenience of understanding and explanation, each logical configuration is described assuming that it is the subject of the operation.

[0059]

[0060] First, the processor (120) can input the generated first learning data (310) into the first feature extractor (230-1). For example, the first feature extractor (230-1) may be a teacher encoder that has been trained in advance to accurately extract features of the input data. Accordingly, the first feature extractor (230-1) can generate information on features of the first learning data (310) (hereinafter, referred to as a 'first feature vector'). In addition, the first feature extractor (230-1) can also generate information on features of each of n data queues (where n is a natural number greater than or equal to 2) pre-stored in the memory (110) (hereinafter, referred to as a 'first data queue feature vector' to an 'n-th data queue feature vector'). The first feature vector, the first data queue feature vector, and the n-th data queue feature vector generated by the first feature extractor (230-1) can be stored in the memory (110).

[0061] The first feature vector, the first data queue feature vector, and the n-th data queue feature vector generated by the first feature extractor (230-1) may be output with the weight set in the first MLP (240-1) applied thereto. Hereinafter, each of the first feature vector, the first data queue feature vector, and the n-th data queue feature vector may be information output from the first MLP (240-1).

[0062]

[0063] Meanwhile, the number of data (n) constituting the data queue can be preset in advance (e.g., 128K). The data queue of a predetermined length can be filled with a random value when the self-supervised learning according to one embodiment of the present invention is initiated, but can be updated in such a way that the data used in the previous iteration is reused each time during the loss function calculation process. For example, assuming that the number of original data used for learning is 500, the length of each data queue is 100, and 25 pieces of original data are grouped into batches and input to the artificial intelligence algorithm in one iteration, the first 25 pieces of original data (d1 to d25) can be input to the artificial intelligence algorithm in the first iteration and used for calculating the similarity distribution together with 100 random data queues (r1 to r100).

[0064] Before proceeding to the second iteration, the 25 batches used in the previous iteration (i.e., the first iteration) can be applied to the data queue, and the data queue can be updated in a pushing manner by removing 25 of the data queues used in the previous iteration. Then, similarity distribution calculation and learning can be performed through the batches of the second iteration (d26 to d50) and the updated data queues (d1 to d25, r26 to r100), and the data queues can be updated again, and this can be repeated.

[0065]

[0066] Additionally, the processor (120) can input the generated second learning data (320) into the second feature extractor (230-2). For example, the second feature extractor (230-2) may be a student encoder that is a target of self-supervised learning. Accordingly, the second feature extractor (230-2) can generate information on the features of the second learning data (320) (hereinafter, referred to as a "second feature vector") before learning is completed.

[0067] The second feature vector generated by the second feature extractor (230-2) may be output with the weight set in the second MLP (240-2) applied. Hereinafter, the second feature vector may be information output from the second MLP (240-2).

[0068]

[0069] In addition, the graph generator (250) can generate 'graph data' based on the first feature vector, the second feature vector, and the first to n-th data queue feature vectors according to a preset method. In one embodiment of the present invention, the 'graph data' may be data expressing the interaction between the first feature vector, the second feature vector, and the first to n-th data queue feature vectors (a total of n+2 data). Therefore, the graph generator (250) can generate a 'node feature matrix' and an 'adjacency matrix' using a total of n+2 data according to a preset method.

[0070] First, the node feature matrix can be generated by the mathematical formula 1 below.

[0071] [Mathematical Formula 1]

[0072]

[0073] That is, the node feature matrix Z is an (n+2)-dimensional matrix, and each dimension can be filled with n+2 data. The node feature matrix can be expressed interchangeably with the 'first feature matrix'.

[0074] In addition, the adjacency matrix A is a matrix with dimensions of (n+2)*(n+2), and can have values ​​of 1 or 0. For example, if A_ij=1, which is the value in the ith row and jth column, it can be interpreted that the ith node (data) and the jth node are 'related'. Conversely, if A_ij=0, it can be interpreted that the ith node (data) and the jth node are not related. The adjacency matrix A can be generated by the following mathematical formula 2.

[0075] [Mathematical Formula 2]

[0076]

[0077] Here, is the transpose of the node feature matrix (Z), and s can be a sigmoid function. The adjacency matrix can be expressed in combination with the second feature matrix.

[0078]

[0079] In addition, the structural feature generator (260-1) can generate structural feature information using the generated graph data (node ​​feature matrix and adjacency matrix). For example, a plurality of MLPs (not shown) having one hidden layer can be stored in the structural feature generator (260-1), and information about the adjacency matrix (A) can be input into the MLP (not shown) to calculate structural feature information based on connectivity. The process of calculating structural feature information in the structural feature generator (260-1) can be largely composed of two processes.

[0080] The first step is a process of updating connectivity-based features, which can be performed using the following mathematical formula 3.

[0081] [Equation 3]

[0082]

[0083] Here, can mean a connectivity-based matrix of (n+2) dimensions corresponding to the node feature matrix. In addition, and may be the weight values ​​of the first MLP (not shown) preset in the structural feature generator (260-1). In addition, may be values ​​corresponding to the node feature matrix in the adjacency matrix. That is, mathematical expression 3 may be a formula that applies the weight of the first MLP (not shown) to all nodes that are connected to an arbitrary node (I).

[0084] Afterwards, the value calculated by mathematical formula 3 can be applied to the second MLP (not shown) as in mathematical formula 4 below to calculate structural feature information.

[0085] [Equation 4]

[0086]

[0087] Here, is a structural feature information, and is a (n+2)-dimensional matrix corresponding to the node feature matrix, and may be weight values ​​of the second MLP (not shown) preset in the structural feature generator (260-1).

[0088] By the above-described method, the structural feature generator (260-1) can generate structural feature information corresponding to the first feature vector and structural feature information corresponding to the second feature vector, respectively. That is, the first feature vector, the second feature vector, and the first to n-th data queue feature vectors (a total of n+2 data) can each be a node, and the method for generating structural feature information corresponding to each node is the above mathematical expressions 1 to 4.

[0089]

[0090] In addition, the node feature generator (260-2) can generate node feature information through a graph neural network model (e.g., GNN layers) based on the node feature matrix Z. For example, the node feature generator (260-2) may include a GNN model, and the node feature matrix Z may be input into the GNN model to output node feature information.

[0091]

[0092] Hereinafter, the structural feature information corresponding to the first feature vector is referred to as the 1-1 feature information. In addition, the node feature information corresponding to the first feature vector is referred to as the 1-2 feature information. In addition, the structural feature information corresponding to the second feature vector is referred to as the 2-1 feature information. In addition, the node feature information corresponding to the second feature vector is referred to as the 2-2 feature information.

[0093] The operation of generating structural feature information in the structural feature generator (260-1) and the operation of generating node feature information in the node feature generator (260-2) are as diagrammed in Fig. 4.

[0094]

[0095] Additionally, structural feature information and node feature information can be combined to create comparative information. For example, feature information 1-1 and feature information 1-2 can be combined to create first comparative information. Furthermore, feature information 2-1 and feature information 2-2 can be combined to create second comparative information.

[0096] In addition, the second feature extractor (230-2) and / or the second MLP (240-2) can be trained using the comparison results of the first and second comparison information. For example, the similarity between the distribution of the first comparison information and the distribution of the second comparison information can be measured through the Kullback-Leibler divergence (270). In addition, the second feature extractor (230-2) and / or the second MLP (240-2) can be trained so that the distribution of the second comparison information follows the distribution of the first comparison information based on the results measured by the Kullback-Leibler divergence (270).

[0097]

[0098] FIG. 5 is a diagram for explaining the effect of an artificial intelligence algorithm learned according to one embodiment of the present invention.

[0099] Referring to FIG. 5, the results of detecting cells in an image using one embodiment of the present invention are illustrated. Class (A) is an image class including basophils, (B) chronic lymphocytic leukemia, (C) reactive lymphocytes, (D) monocytes, and (E) splenic marginal zone lymphoma.

[0100] The blue and red boxes depicted in Figure 5 represent the bounding boxes for actual and predicted values, respectively. This demonstrates that the present invention can automatically detect cells with great precision.

[0101] Therefore, it is clear that the self-supervised learning device (100) of the artificial intelligence algorithm according to the present invention can train the artificial intelligence algorithm by considering not only the direct connectivity (first-order connectivity) between the original data (300) and the data queue, but also the connectivity (higher-order connectivity) between the data queues, thereby improving accuracy.

[0102]

[0103] The self-supervised learning operation of the artificial intelligence algorithm according to the present invention described above can be implemented as a computer-readable code in a memory (110), which is a computer-readable recording medium. The computer-readable recording medium includes all types of recording media that store data that can be deciphered by a computer system. For example, there may be a ROM (Read Only Memory), a RAM (Random Access Memory), a magnetic tape, a magnetic disk, a flash memory, an optical data storage device, etc. In addition, the computer-readable recording medium can be distributed to computer systems connected to a computer communication network, and stored and executed as a code that can be read in a distributed manner. Accordingly, the processor (120) can execute the program instructions (code) stored in the memory (110) to perform the text recognition operation of the target image described above.

[0104]

[0105] Above, the present invention has been described in detail with reference to preferred embodiments, but the present invention is not limited to the above embodiments, and various modifications and changes are possible by those skilled in the art within the technical spirit and scope of the present invention.

Claims

1. Memory where original data and artificial intelligence algorithms are stored; and processor; Including, but not limited to, The above memory is executable by the processor, A self-supervised learning device that inputs learning data corresponding to the original data into the first feature extractor and the second feature extractor of the artificial intelligence algorithm, generates 1-1 feature information and 1-2 feature information using the result of the first feature extractor, generates 2-1 feature information and 2-2 feature information using the result of the second feature extractor, and stores program commands for training the second feature extractor using the 1-1 feature information, the 1-2 feature information, the 2-1 feature information, and the 2-2 feature information.

2. In paragraph 1, The above memory is executable by the processor, A self-supervised learning device that stores program commands for generating first learning data and second learning data using the original data, inputting the first learning data into the first feature extractor, and inputting the second learning data into the second feature extractor.

3. In paragraph 2, The above memory is executable by the processor, A self-supervised learning device that stores program commands for generating at least one of the first learning data and the second learning data through at least one of color change, rotation, and partial selection of the original data.

4. In paragraph 1, The above memory further comprises n data queues, executable by the processor, Store program commands for generating the 1-1 feature information, the 1-2 feature information, the 2-1 feature information, and the 2-2 feature information using the 1st feature vector, the 2nd feature vector, and the feature vectors of each of the n data queues, The above first feature vector is the output value of the first feature extractor, The above second feature vector is the output value of the second feature extractor, The feature vector of each of the above n data queues is a preset value or information generated in advance through the first feature extractor, A self-supervised learning device, wherein n is a natural number greater than or equal to 2.

5. In paragraph 4, The above memory is executable by the processor, A self-supervised learning device storing program commands that produce the above-mentioned first-first feature information by mathematical expression 1 and mathematical expression 4. [Mathematical Formula 1] [Mathematical Formula 2] [Equation 3] [Equation 4] However, the above Z is a first feature matrix of (n+2) dimensions, and is the transpose matrix of the first feature matrix, A is an adjacency matrix of dimension (n+2)*(n+2), s is a sigmoid function, and is a connectivity-based matrix of (n+2) dimensions corresponding to the first feature vector, and and above are the weight values ​​of the first MLP stored in advance, and the above are the values ​​corresponding to the first feature vector in the above adjacency matrix, and is the above 1-1 feature information, and is a (n+2)-dimensional matrix corresponding to the first feature vector, and and above are the weight values ​​of the second MLP that were previously stored.

6. In paragraph 5, The above memory is executable by the processor, Store program commands that produce the 2-1 characteristic information by the above mathematical formula 1 and the above mathematical formula 4, Above is a connectivity-based matrix of (n+2) dimensions corresponding to the second feature vector, and and above are the weight values ​​of the first MLP stored in advance, and the above are the values ​​corresponding to the second feature vector in the above adjacency matrix, and is the above 2-1 feature information, and is a (n+2)-dimensional matrix corresponding to the second feature vector, and and above are the weight values ​​of the second MLP that were previously stored.

7. In paragraph 6, The above memory is executable by the processor, A self-supervised learning device that stores program commands for producing the first feature matrix using the first-second feature information and the second-second feature information.

8. In paragraph 7, The above memory is executable by the processor, A self-supervised learning device storing program commands for generating first comparison information using the first-1 feature information and the first-2 feature information, generating second comparison information using the second-1 feature information and the second-2 feature information, and training the second feature extractor by comparing the first comparison information and the second comparison information.

9. In paragraph 8, A self-supervised learning device, wherein the first comparison information is a sum of the first-1 feature information and the first-2 feature information, and the second comparison information is a sum of the second-1 feature information and the second-2 feature information.

10. In paragraph 8, The above memory is executable by the processor, A self-supervised learning device that stores program commands for training the second feature extractor so that the similarity distribution of the second comparison information follows the similarity distribution of the first comparison information.

Images