Method and apparatus for semi-supervised online learning

Abstract

We provide an online semi-supervised learning method and device based on Label Propagation Adaptive Resonance Theory. This enables the training of artificial intelligence models using small amounts of labeled data continuously in real-world usage environments.

Description

Technology Field

[0001] The present invention relates to an online semi-supervised learning method and apparatus, and more specifically, to an online semi-supervised learning method and apparatus based on Label Propagation Adaptive Resonance Theory. Background Technology

[0002] The contents described below are provided solely for the purpose of providing background information related to embodiments of the present invention, and the contents described do not automatically constitute prior art.

[0003] Over the past few years, deep neural network models have shown remarkable progress, particularly in visual object recognition, speech recognition, and autonomous robot control. However, deep learning has a practical drawback: it requires not only the collection of massive amounts of data but also time and effort for labeling. In this regard, the fields of semi-supervised learning, continuous learning, transfer learning, and meta-learning have recently been gaining attention.

[0004] Semi-supervised learning addresses problems where unlabeled data is abundant but labeled data is extremely limited.

[0005] Online continuous learning aims to enable models to adapt to a constantly changing environment by learning from sequential data without forgetting previous knowledge.

[0006] Meanwhile, although various label propagation methods for semi-supervised learning have been studied, most methods require the repeated use of the same data for training and are not suitable for environments where opportunities to repeatedly access the same data are limited.

[0007] Although the above paradigms have each shown remarkable research results, they show limitations in real-world problems where labeled data is rarely provided, such as in household robots and medical data, and where repeated access to data learned once is restricted due to privacy issues.

[0008] The aforementioned prior art refers to technical information that the inventor possessed for the derivation of the present invention or acquired during the process of deriving the present invention, and it cannot be considered as publicly known technology disclosed to the general public prior to the filing of the present invention. The problem to be solved

[0009] One aspect of the embodiment is to provide an online semi-supervised learning method and apparatus based on online label propagation suitable for continuous learning of an ART-based network.

[0010] One aspect of the embodiment is to provide an online semi-supervised learning method and apparatus capable of continuous learning without the need to reuse data that has been learned once.

[0011] One aspect of the embodiment is to provide an online semi-supervised learning method and apparatus that provides uncertainty regarding label inference results.

[0012] The objectives of the present invention are not limited to the problems mentioned above, and other objectives and advantages of the present invention not mentioned may be understood from the following description and will be more clearly understood from the embodiments of the present invention. Furthermore, it will be understood that the objectives and advantages of the present invention can be realized by the means and combinations thereof set forth in the claims. means of solving the problem

[0013] An online semi-supervised learning method according to one embodiment may include the steps of generating a set of activation nodes for input data among at least one output node of a clustering network, performing label propagation on activation nodes belonging to the set of activation nodes, and reflecting the input data to the maximum activation node of the set of activation nodes.

[0014] An online semi-supervised learning device according to one embodiment includes a memory for storing at least one instruction and a processor, wherein the at least one instruction may be configured such that when executed by the processor, the processor generates a set of activation nodes for input data among at least one output node of a clustering network, performs label propagation for activation nodes belonging to the set of activation nodes, and reflects the input data to the maximum activation node of the set of activation nodes.

[0015] Other aspects, features, and advantages other than those described above will become apparent from the following drawings, claims, and detailed description of the invention. Effects of the invention

[0016] According to the embodiment, it is possible to train an artificial intelligence model using a small amount of labeled data continuously in an actual usage environment.

[0017] According to the embodiment, the need to pre-create a large amount of labeled data to apply artificial intelligence to industries or environments is alleviated.

[0018] According to the embodiment, even in situations where there are extremely few labels, the recognition performance of machine learning can be improved by utilizing unlabeled data.

[0019] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art from the description below. Brief explanation of the drawing

[0020] FIG. 1 is a diagram for schematically illustrating online semi-supervised learning according to an embodiment. FIG. 2 is a block diagram of an online semi-supervised learning device according to an embodiment. Figure 3 is a flowchart of an online semi-supervised learning method according to an embodiment. FIG. 4 is a diagram illustrating an online semi-supervised learning process according to an embodiment. FIG. 5 shows exemplary pseudocode of an online semi-supervised learning algorithm according to an embodiment. Figure 6 is a graph showing experimental results by online semi-supervised learning according to an embodiment. Specific details for implementing the invention

[0021] The present invention will be described in more detail below with reference to the drawings. The present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the following embodiments, parts not directly related to the description are omitted to clearly explain the present invention; however, this does not mean that such omitted components are unnecessary when implementing a device or system to which the concept of the present invention is applied. Furthermore, the same reference numerals are used for identical or similar components throughout the specification.

[0022] In the following description, terms such as "first," "second," etc., may be used to describe various components, but said components should not be limited by said terms, and said terms are used solely for the purpose of distinguishing one component from another. Additionally, in the following description, singular expressions include plural expressions unless the context clearly indicates otherwise.

[0023] In the following description, terms such as “comprising” or “having” are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not excluding in advance the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0024] Most commercially successful AI algorithms rely on models pre-trained on large volumes of labeled data. However, performance degradation frequently occurs in diverse real-world environments due to differences from the pre-trained environment. Furthermore, most label propagation methods require the repeated use of the same data for training, making them unsuitable for environments where repeated access to identical data is limited.

[0025] Meanwhile, Adaptive Resonance Theory (ART) is a solution for online continuous learning inspired by the brain's information processing mechanisms.

[0026] The present invention proposes a semi-supervised online learning method for applying machine learning to such real-world problems. Specifically, the embodiment proposes an online semi-supervised learning method and apparatus that provide online label propagation suitable for continuous learning based on an ART network.

[0027] According to an embodiment, a machine learning algorithm capable of continuous online learning even with rarely labeled data is provided.

[0028] The following describes embodiments in detail with reference to the drawings.

[0029] FIG. 1 is a diagram for schematically illustrating online semi-supervised learning according to an embodiment.

[0030] An online semi-supervised learning device (100) according to an embodiment executes an online semi-supervised learning method. The specific configuration of the online semi-supervised learning device (100) will be described later with reference to FIG. 2.

[0031] An online semi-supervised learning device (100) according to an embodiment processes original input data (IN_DATA) one by one from a stream of original input data (IN_DATA), which is a series of sequential data.

[0032] An online semi-supervised learning device (100) according to an embodiment outputs label information for raw data (xi) from original input data (IN_DATA), that is, label information for input data (ri) to be described later. For example, the label information includes label density (q) and uncertainty (u).

[0033] In one example, the original input data (IN_DATA) may include raw data (xi) for images, sound, voice, and text. In one example, the original input data (IN_DATA) includes label information (y i ) may be provided together or label information may be missing.

[0034] An online semi-supervised learning device (100) can output a label density (q) for raw data (xi) by executing a classification network and an online label propagation algorithm. Additionally, the online semi-supervised learning device (100) can output at least one uncertainty (u) for the label density (q) by an uncertainty determination algorithm.

[0035] FIG. 2 is a block diagram of an online semi-supervised learning device according to an embodiment.

[0036] An online semi-supervised learning device (100) according to an embodiment includes a processor (110) and a memory (120).

[0037] The processor (110) is a type of central processing unit and can control the operation of the online semi-supervised learning device (100) by executing one or more instructions stored in memory (120).

[0038] The processor (110) includes all kinds of devices capable of processing data. The processor (110) may mean a data processing device embedded in hardware, having a physically structured circuit to perform a function expressed by code or instructions included in a program, for example.

[0039] Examples of data processing devices embedded in hardware as described above include, but are not limited to, processing devices such as a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), and a graphics processing unit (GPU). The processor (110) may include one or more processors.

[0040] The memory (120) is operably connected to the processor (110) and stores at least one instruction, i.e., at least one executable code, that is, is executed by the processor (110).

[0041] The memory (120) may store a program containing one or more instructions for processing original input data (IN_DATA) and outputting label information. The processor (110) may execute an online semi-supervised learning process according to the embodiment based on the program and instructions stored in the memory (120).

[0042] The memory (120) can store a computer program that implements classification network information and an algorithm for label propagation for implementing online semi-supervised learning according to the embodiment.

[0043] The memory (120) can further store intermediate data and computation results, etc., generated during the execution of the online semi-supervised learning method according to the embodiment. The memory (120) stores a feature vector (r) from raw data (xi). i Encoder and feature vector (r) that extract ) i Training data, intermediate data, and result data of a classification network (Clusterint Network) that classifies ) can be stored.

[0044] The memory (120) may include internal memory and / or external memory, and may include volatile memory such as DRAM, SRAM, or SDRAM, non-volatile memory such as OTPROM (One Time Programmable Read Only Memory), PROM (Programmable Read Only Memory), EPROM (Erasable PROM), EEPROM (Electrically Erasable PROM), mask ROM, flash ROM, NAND flash memory, or NOR flash memory, flash drives such as SSD (Solid State Drive), CF (Compact Flash) card, SD card, Micro-SD card, Mini-SD card, Xd card, or memory stick, or storage devices such as HDD (Hard Disk Drive). The memory (120) may include magnetic storage media or flash storage media, but is not limited thereto.

[0045] Additionally, the online semi-supervised learning device (100) may further include a communication unit (not shown) that performs data transmission and reception with an external device.

[0046] For example, an online semi-supervised learning device (100) may receive information for the execution of a model, such as initial weights, hyperparameters, and a loss function of a feature encoder and / or classification network, through a communication unit and store this information in memory (120). For example, the online semi-supervised learning device (100) may receive at least one feature vector (r) extracted from original input data (IN_DATA) or raw data (xi) of original input data (IN_DATA) from an external device through a communication unit. i Can receive ).

[0047] The online semi-supervised learning device (100) executes an online semi-supervised learning method according to an embodiment. That is, the online semi-supervised learning device (100) according to an embodiment outputs label information from original input data (IN_DATA).

[0048] The operation of an online semi-supervised learning device (100) for executing an online semi-supervised learning method according to an embodiment is described schematically below, and the specific operation is described later with reference to FIG. 4.

[0049] At least one instruction stored in memory (120) can be configured to cause the processor (110), when executed by the processor (110), to generate a set of activation nodes for input data (ri) among at least one output node of the clustering network, perform label propagation on the activation nodes belonging to the set of activation nodes, and reflect the input data (ri) to the maximum activation node of the set of activation nodes. Here, the input data (ri) includes a feature vector of raw data (xi).

[0050] In one example, the classification network includes a learning model based on Adaptive Resonance Theory (ART). In one example, the output node of the classification network can output the label density (q) of the input data (ri) classified by the output node according to the label density function of the output node.

[0051] The input data (ri) includes a feature vector of the original input data (IN_DATA), and when at least one instruction stored in memory (120) is executed by the processor (110), the processor (110) generates a second feature vector (I) based on the complement vector of the input data (ri) for each output node of at least one output node in order to generate a set of activation nodes. i It can be configured to compute the comparison function and selection function of the classification network based on the profile parameters of each output node. Here, the second feature vector (Ii) is a vector formed by concatenating the input data (ri) and the complement vector of the input data (ri).

[0052] At least one instruction stored in memory (120) can be configured to cause the processor (110) to add an output node whose operation result of a comparison function is greater than or equal to a threshold value to the set of active nodes as an active node when executed by the processor (110) to create a set of active nodes.

[0053] At least one instruction stored in memory (120) causes the processor (110) to label (y) the input data (ri) to generate a set of activation nodes when executed by the processor (110). i Given ), the comparison function can be configured to update the label density function of the output node whose operation result is greater than or equal to the threshold.

[0054] At least one instruction stored in memory (120) can be configured to update the label density function of each activation node based on the label density function of all activation nodes belonging to the set of activation nodes when executed by the processor (110) in order to perform label propagation.

[0055] When at least one instruction stored in memory (120) is executed by the processor (110), the processor (110) determines the activation node with the maximum result of the selection function as the maximum activation node in order to reflect the input data, and the second feature vector (I) of the input data (ri). i It can be configured to update the profile parameters of the maximum active node based on the profile parameters of the maximum active node and the profile parameters of the maximum active node.

[0056] At least one instruction stored in memory (120) can be configured to cause the processor (110) to determine the label uncertainty of the original input data (IN_DATA) based on the label density function of the maximum active node when executed by the processor (110).

[0057] Figure 3 is a flowchart of an online semi-supervised learning method according to an embodiment.

[0058] In addition to the steps (S2) to (S4) to be described later, the online semi-supervised learning method according to the embodiment executes a pre-trained encoder based on a series of original input data (IN_DATA) by the processor (110) prior to step (S2) to the series of input data (r i ), that is, a step (S1) of obtaining a series of feature vectors may be further included. For example, a variational autoencoder (VAE) may be used as the previously trained encoder, and this will be described later with reference to FIG. 4.

[0059] An online semi-supervised learning method according to an embodiment may include the steps of: generating a set of activation nodes for input data among at least one output node of a clustering network by a processor (110) (S2); performing label propagation on activation nodes belonging to the set of activation nodes (S3); and reflecting input data (ri) to the maximum activation node of the set of activation nodes (S4).

[0060] In step (S2), the processor (110) may generate a set of activation nodes (A with reference to FIG. 5) for the input data among at least one output node of the clustering network. In one example, the clustering network may be a learning model based on Adaptive Resonance Theory (ART). In one example, the output node of the clustering network may be configured to output the label density of the input data classified to the output node according to the label density function of the output node.

[0061] In one example, the input data (ri) includes a feature vector of the original input data, and step (S2) is performed by the processor (110) for each output node of at least one output node, the input data (r i The method may include the step of calculating a comparison function (Vj with reference to FIG. 5) and a selection function (Tj with reference to FIG. 5) of a classification network based on a second feature vector (Ii) based on the complement vector of ) and profile parameters of each output node.

[0062] In step (S2), the processor (110) can calculate a comparison function (Vj with reference to FIG. 5) and a selection function (Tj with reference to FIG. 5) according to the following mathematical formula.

[0063]

[0064] Here, Ii is the second feature vector, and ∧ represents the pairwise minimum. α is a constant, and wj represents the profile parameter of node j.

[0065] The node's profile parameter w is the feature vector of that node, and its vector value is determined based on the input data classified as that node so far.

[0066] Step (S2) may further include the step of adding an output node whose operation result of a comparison function (Vj) is greater than or equal to a threshold value to the set of activated nodes (A) as an activated node by the processor (110).

[0067] Step (S2) may further include, when a label is given to the input data (ri), updating the label density function (qj) of output node j by the processor (110) such that the result of the operation of the comparison function (Vj) is greater than or equal to a threshold (vigilance parameter ρ with reference to FIG. 6). For example, the label density function (qj) may be increased by 1 for the label (yi) of the input data (see lines 9-10 of FIG. 6).

[0068] Subsequently, in step (S3), the processor (110) performs label propagation for the activation nodes belonging to the set of activation nodes (A).

[0069] Step (S3) may include updating the label density function of each activation node based on the label density function (q) of all activation nodes belonging to the set of activation nodes (A) by the processor (110).

[0070] Here, according to the following mathematical formula, for all nodes k belonging to the set of activation nodes (A), if there are no actually collected labels for the corresponding nodes, an update of the label density function (q) can be performed for the nodes k.

[0071]

[0072] Here, delta (δ) is the propagation rate constant, A is the set of all nodes that are activated together, and k is the node currently receiving propagation. The constant C is used to ensure that the magnitude of the label density function formed by propagation does not exceed 1.

[0073] In step (S4), the processor (110) can reflect input data (ri) to the maximum active node (J) of the set of active nodes (A).

[0074] In one example, step (S4) is a step in which the processor (110) determines the activation node whose operation result of the selection function (Tj) described in step (S2) is maximum as the maximum activation node (J), and input data (r i The second feature vector (I) of ) i It may include a step of updating the profile parameters of the maximum active node based on the profile parameters of the maximum active node and the maximum active node.

[0075] For example, the processor (110) can update the profile parameters of the maximum active node (J) according to the following mathematical formula 3.

[0076]

[0077] Here, wJ is the profile parameter of the maximum active node (J), and β represents the learning rate.

[0078] The online semi-supervised learning method according to the embodiment may further include a step (S5) in which the processor (110) determines the label uncertainty of the input data (ri) based on the label density function of the maximum active node (J). For example, in step (S5), the processor (110) may determine the label uncertainty based on Equations 4 and 5, which will be described later with reference to FIG. 4.

[0079] Meanwhile, the online semi-supervised learning method according to the embodiment may further include the step of executing the steps (S1) to (S5) described above on the next original input data that is sequentially input.

[0080] Meanwhile, if there are no activation nodes based on the input data (ri) as a result of executing step (S2), that is, if the set of activation nodes (A) is empty, the processor (110) creates a new output node for the original input data (IN_DATA) (corresponding to lines 16-17 with reference to FIG. 5). In this case, the processor (110) sets the label density (qn(yi)) for the label (yi) of the original input data (IN_DATA) of the newly created node to 1.

[0081] Steps (S1) to (S5) will be examined in detail below with reference to FIGS. 4 and FIGS. 5.

[0082] FIG. 4 is a diagram illustrating an online semi-supervised learning process according to an embodiment.

[0083] Feature extraction corresponds to step (S1) with reference to FIG. 3.

[0084] For example, in step (S1), the processor (110) performs VAE on high-dimensional raw data (xi) to encode output vectors (σ and μ), and among these, a low-dimensional feature vector (μ) corresponding to at least a portion of the output vector (μ). d Extract ) as input data (ri).

[0085] Most raw data (xi), such as images and sounds, exists in the form of high-dimensional original data, and in step (S1), the processor (110) can represent it as a low-dimensional feature vector (ri) so that the ART network can process it appropriately. For this purpose, for example, a Variational Autoencoder (VAE) was used as a feature extractor, but various feature extraction algorithms can be utilized.

[0086] The Fuzzy ART network corresponds to steps (S2) and (S4) with reference to FIG. 3. For example, the classification network is illustrated as a Fuzzy ART network, but is not limited thereto.

[0087] ART-based classification networks are self-organizing neural networks inspired by brain information processing mechanisms (Stephen Grossberg and Gail Carpenter, 1976) that learn adaptively using the interaction between 'top-down' expectations of input data (ri) and 'bottom-up' sensory information.

[0088] For example, when input data (ri) is input, the fuzzy ART network updates the existing output node (category) if there is a sufficiently similar one among the existing output nodes (categories) (Line 15, described later with reference to FIG. 5), and if there is no similar one, creates a new output node (category) (Line 17, described later with reference to FIG. 5).

[0089] This method is suitable for continuous online learning because it clusters input data (ri) online, updates the label density function (q) in the corresponding output node (cluster) when a label (yi) is input, and allows learning new data while preserving existing learning results.

[0090] Online label propagation corresponds to step (S3) with reference to FIG. 3. When input data (ri) is input to the aforementioned fuzzy ART network, if two or more of the existing learned output nodes are activated, information (labels) are exchanged between the nodes based on the aforementioned mathematical formula 2.

[0091] If this process occurs repeatedly, even output nodes that previously did not have labels assigned can infer labels by utilizing the topological relationships of neighboring output nodes.

[0092] This assumes that when a single input data (ri) activates two or more nodes, the nodes that are activated together are likely to have the same label, and when multiple labels are mixed in a single node, the label with the highest density can be identified.

[0093] For example, in the example of FIG. 4, it is assumed that the original input data (IN_DATA) is a bird image, and two output nodes (q1 and q2) out of a total of four output nodes (q1 to q4) are activated by the input data (ri) for the bird image. In the example, the label density of the new node is updated by performing label propagation according to Equation 2 from the label density of the two output nodes (q1 and q2).

[0094] Next, the uncertainty estimation corresponds to step (S5) with reference to FIG. 3.

[0095] When input data (ri) is input, the label of the input data (ri) is inferred based on the label density information of the node with the highest activation level in the aforementioned step (S4). At this time, various labels may be mixed in the label density information, or the label density itself may be low; in such cases, the uncertainty can be inferred by the following mathematical formula, and the reliability of the inference result can be evaluated.

[0096]

[0097]

[0098] Here, J represents the most active node for the input data (ri), and P J represents the label probability of the corresponding node. q J represents the label density function collected and propagated by the input labels, where P J It is calculated as shown in the mathematical formula below.

[0099]

[0100] The first uncertainty (u 1 ) and the second uncertainty (u 2 ) each represent a different level of uncertainty; first, the first uncertainty (u 1 ) corresponds to the entropy that measures how impurely the labels of the corresponding node are mixed. The second uncertainty (u 2 ) represents the uncertainty regarding the sum of the total label density functions and indicates how many labels have been collected and propagated to the corresponding node.

[0101] Nodes with few collected or unpropagated labels have the second uncertainty (u 2 ) appears high. The first uncertainty (u 1 ) and the second uncertainty (u 2 You can filter out results with low reliability by utilizing at least one of the following.

[0102] FIG. 5 shows exemplary pseudocode of an online semi-supervised learning algorithm according to an embodiment.

[0103] First, the high-dimensional raw data (xi) is converted into input data (ri), which is a low-dimensional feature vector (line2).

[0104] For the application of a classification network, for example, the Fuzzy ART algorithm, a second feature vector (Ii) is generated by concatenating the feature vector of the input data (ri) and its complement vector (line3).

[0105] Nodes that are activated in the classification network are searched based on the second feature vector (Ii) (lines 5 to 11). At this time, if there is a label (yi) in the input data (ri), the label density function (q) of the activated node is updated (lines 9 to 10).

[0106] If there are multiple active nodes, labels are propagated among them (lines 12 to 13). The maximum active node J is found (line 14), and the profile parameter w of that node is updated to reflect the new data in the existing maximum active node (line 15). If there are no active nodes, a new node is created based on the input data (lines 16 to 17).

[0107] Line 2 corresponds to step (S1) with reference to FIG. 3, lines 3 to 11 correspond to step (S2), lines 13 and 17 correspond to step (S3), and lines 14 to 15 correspond to step (S4).

[0108] Figure 6 is a graph showing experimental results by online semi-supervised learning according to an embodiment.

[0109] Figure 6(a) shows the experimental results using the NSynth dataset (instrument sound classification) of the method proposed in the present invention (LPART) and the existing method (FAM).

[0110] The proposed method (LPART) demonstrates a rapid increase in recognition performance compared to the existing supervised learning method (FAM) even with a small amount of data (epochs) and labels, and shows high performance when used with unlabeled data.

[0111] Figure 6 (b) shows the results according to the uncertainty (reliability) of the input data inference results.

[0112] The first uncertainty (u 1 ) and the second uncertainty (u 2 A threshold was set for each, and if the uncertainty was higher than a specific threshold, it was classified as an unreliable result.

[0113] At this time, the accuracy of total samples shows a recognition success rate of about 90 percent, while the accuracy of reliable samples shows a consistently high accuracy of about 98 percent. As training continues, the number of unreliable input samples decreases as labels are gradually collected.

[0114] A method according to one embodiment of the present invention described above can be implemented as computer-readable code on a medium on which a program is recorded. A computer-readable medium is a non-transient recording medium and includes all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SSD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

[0115] Although artificial intelligence has recently shown remarkable progress, it requires a large amount of labeled training data. This can be seen as a major factor slowing down the ripple effect of AI's application in industry.

[0116] The present invention reduces the need to secure large amounts of labeled data in advance to apply artificial intelligence to industry. Furthermore, it enables the application of machine learning in environments where labels are rarely provided and data storage and repetitive access are restricted due to privacy issues. In other words, it enables the application of artificial intelligence to various industrial sectors where it was previously impossible to apply AI due to the relative difficulty of data collection and labeling.

[0117] In addition, the uncertainty measurement method proposed in this invention enables the conservative use of artificial intelligence based on the reliability of results, even in industrial sectors where the production of incorrect results by artificial intelligence can be fatal. This perspective helps to apply artificial intelligence not only to existing business areas but also to various business areas where data labeling is difficult and the use of personal information is difficult.

[0118] 1) Household AI such as robot vacuums and home appliances - Household electronic devices often have restrictions on data storage and transmission to servers due to personal information issues. In such cases, AI can be trained using collected data with prior consent, but performance degradation may occur due to differences in each home environment.

[0119] The method proposed in the present invention mitigates such performance degradation and enables the provision of a solution optimized for each environment.

[0120] 2) Medical Artificial Intelligence - Medical artificial intelligence faces difficulties in storing and transmitting data due to personal information issues, as well as in acquiring large amounts of labeled data due to high labeling (diagnosis) costs.

[0121] In such cases, to create a model that predicts disease based on patient data, unsupervised learning is performed by continuously inputting input data into the invention model, and by inputting labeling (diagnosis result input) for the corresponding data when a doctor's diagnosis is available, the accuracy of the model can be gradually increased without a large amount of training data.

[0122] Particularly in situations where errors in diagnostic results can be fatal, the uncertainty measurement method proposed in this invention allows for assessing the reliability of the results and utilizing them conservatively.

[0123] The description of the embodiments of the present invention described above is for illustrative purposes only, and those skilled in the art will understand that other specific forms can be easily modified without altering the technical spirit or essential features of the present invention. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. For example, each component described as a single unit may be implemented in a distributed manner, and components described as distributed may likewise be implemented in a combined form.

[0124] The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and equivalent concepts thereof should be interpreted as being included within the scope of the present invention. Explanation of the symbols

[0126] 100: Online semi-supervised learning device 110: Processor 120: Memory

Claims

Claim 1 An online semi-supervised learning method performed by an online semi-supervised learning device, comprising: a step of generating a set of activation nodes for input data among at least one output node of a clustering network; a step of performing label propagation on activation nodes belonging to the set of activation nodes; and a step of reflecting the input data to the maximum activation node of the set of activation nodes, wherein the input data includes a feature vector of original input data, and the step of generating the set of activation nodes further comprises, for each output node of the at least one output node, a step of computing a comparison function and a selection function of the clustering network based on a second feature vector based on the complement vector of the feature vector and a profile parameter of each output node, and a step of determining the label uncertainty of the input data based on the label density function of the maximum activation node, wherein the label uncertainty of the input data includes a first uncertainty corresponding to the entropy of the impurity of the label of the corresponding node and a second uncertainty corresponding to the entropy of the sum of the total label density functions. Claim 2 In claim 1, the classification network is an online semi-supervised learning method, which is a learning model based on Adaptive Resonance Theory (ART). Claim 3 An online semi-supervised learning method according to claim 1, wherein the output node of the classification network outputs the label density of the input data classified by the output node according to the label density function of the output node. Claim 4 delete Claim 5 An online semi-supervised learning method according to claim 1, wherein the step of generating the set of activation nodes further includes the step of adding an output node whose operation result of the comparison function is greater than or equal to a threshold to the set of activation nodes as the activation node. Claim 6 An online semi-supervised learning method according to claim 5, wherein the step of generating the set of activation nodes further includes the step of updating the label density function of an output node whose operation result of the comparison function is greater than or equal to a threshold when a label is given to the input data. Claim 7 An online semi-supervised learning method according to claim 1, wherein the step of performing label propagation includes the step of updating the label density function of each activation node based on the label density function of all activation nodes belonging to the set of activation nodes. Claim 8 An online semi-supervised learning method according to claim 1, wherein the step of reflecting the input data comprises: determining the activation node whose operation result value of the selection function is maximum as the maximum activation node; and updating the profile parameter of the maximum activation node based on the second feature vector and the profile parameter of the maximum activation node. Claim 9 delete Claim 10 An online semi-supervised learning method according to claim 1, further comprising the step of obtaining a series of input data by executing a pre-trained encoder based on a series of raw input data. Claim 11 An online semi-supervised learning device comprises: a memory for storing at least one instruction; and a processor, wherein when the at least one instruction is executed by the processor, the processor generates a set of activation nodes for input data among at least one output node of a clustering network, performs label propagation on activation nodes belonging to the set of activation nodes, reflects the input data to the maximum activation node of the set of activation nodes, the input data includes a feature vector of the original input data, and to generate the set of activation nodes, for each output node of the at least one output node, computes a comparison function and a selection function of the clustering network based on a second feature vector based on the complement vector of the feature vector and a profile parameter of each output node, and determines the label uncertainty of the input data based on the label density function of the maximum activation node, wherein the label uncertainty of the input data includes a first uncertainty corresponding to the entropy of the impurity of the label of the corresponding node and a second uncertainty corresponding to the entropy of the sum of the total label density functions. Claim 12 In claim 11, the classification network is an online semi-supervised learning device that is a learning model based on Adaptive Resonance Theory (ART). Claim 13 An online semi-supervised learning device according to claim 11, wherein the output node of the classification network outputs the label density of the input data classified by the output node according to the label density function of the output node. Claim 14 delete Claim 15 An online semi-supervised learning device according to claim 11, wherein the at least one instruction is configured to cause the processor, when executed by the processor, to add an output node whose operation result of the comparison function is greater than or equal to a threshold value to the set of activation nodes as the activation node in order to generate the set of activation nodes. Claim 16 An online semi-supervised learning device according to claim 15, wherein the at least one instruction is configured to cause the processor, when executed by the processor, to update the label density function of an output node whose operation result of the comparison function is greater than or equal to a threshold when the input data is labeled in order to generate the set of activation nodes. Claim 17 An online semi-supervised learning device according to claim 11, wherein the at least one instruction is configured to enable the processor, when executed by the processor, to update the label density function of each activation node based on the label density function of all activation nodes belonging to the set of activation nodes in order to perform the label propagation. Claim 18 An online semi-supervised learning device according to claim 11, wherein the at least one instruction is configured to cause the processor, when executed by the processor, to determine the activation node having the maximum result of the operation of the selection function as the maximum activation node in order to reflect the input data, and to update the profile parameter of the maximum activation node based on the second feature vector and the profile parameter of the maximum activation node. Claim 19 delete Claim 20 A computer-readable, non-transient recording medium storing a computer program comprising at least one instruction configured to execute an online semi-supervised learning method according to any one of claims 1 to 3, 5 to 8, and 10 by a processor.

Images