Electronic device for training artificial intelligence model, and training method therefor
Training AI models using a similarity map in a Top-K method addresses the limitations of conventional Top-1 classification by enhancing the representation of similar objects, improving accuracy and adaptability.
Patent Information
- Application Number
- PCT/KR2026/002986
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-16
- Filing Date
- 2026-02-23
- Publication Date
- 2026-11-19
AI Technical Summary
Conventional AI models for image data classification are limited in training performance as they focus solely on the correct class, failing to adequately train for classes other than the correct one, leading to inadequate representation of similar objects.
The use of a similarity map to train AI models in a Top-K manner, which considers multiple candidates and incorporates feedback on feature vectors of top K objects to adjust model parameters based on similarity relationships.
This approach enhances the flexibility and adaptability of AI models to real-world environments by accurately identifying and incorporating semantically similar objects, improving accuracy and responsiveness to user intent.
Smart Images

Figure KR2026002986_19112026_PF_FP_ABST
Abstract
Description
Electronic device for training an artificial intelligence model and the training method thereof
[0001] The present disclosure relates to an electronic device for training an artificial intelligence model and a method for training the same.
[0002] As artificial intelligence technology develops, AI models are being used for various purposes in diverse fields. One example is the use of AI models for image data classification.
[0003] Specifically, when specific content is input, the AI model can identify and inform what object the image included in that content represents.
[0004] In the case of artificial intelligence models for image data classification, it was conventionally common practice to train them based on Top-1 classification. In this approach, while the model learns well for the correct class, it fails to adequately train for classes other than the correct one, leading to a problem where the model's training performance is limited.
[0005] An electronic device according to at least one embodiment of the present disclosure includes a memory storing a map corresponding to similarity between a plurality of objects and at least one processor, wherein the at least one processor acquires the correct object and at least one other object having a lower similarity to the object than the correct object based on output data acquired through an artificial intelligence model in relation to input data containing the object, and controls the adjustment of at least one model parameter of the artificial intelligence model to reduce the learning loss associated with the at least one other object based on the difference between the similarity between the correct object and the at least one other object acquired through the artificial intelligence model and the similarity recorded in the map.
[0006] A learning method for an electronic device for learning an artificial intelligence model according to one embodiment of the present disclosure comprises: acquiring output data output from the artificial intelligence model in relation to input data containing an object; acquiring a correct object corresponding to the object and at least one other object having a lower similarity to the object than the correct object based on the output data; identifying the similarity between the correct object and the at least one other object acquired through the artificial intelligence model based on a map corresponding to similarities between a plurality of objects; and controlling at least one model parameter of the artificial intelligence model to adjust the learning loss associated with the at least one other object based on the difference between the similarity between the correct object and the at least one other object acquired through the artificial intelligence model and the similarity recorded in the map containing the object.
[0007] A non-transient computer-readable recording medium storing a program for executing a learning method of an electronic device for learning an artificial intelligence model according to one embodiment of the present disclosure comprises: a step of obtaining output data output from the artificial intelligence model in relation to input data containing an object; a step of obtaining the correct object and at least one other object having a lower similarity to the object than the correct object based on the output data; a step of identifying the similarity between the correct object obtained through the artificial intelligence model and the at least one other object based on a map corresponding to similarities between a plurality of objects; and a step of controlling at least one model parameter of the artificial intelligence model to adjust the learning loss for the at least one other object based on the difference between the similarity between the correct object obtained through the artificial intelligence model and the at least one other object and the similarity recorded in the map containing the object.
[0008] FIG. 1 is a drawing for explaining the operation of an electronic device according to one or more embodiments of the present disclosure.
[0009] FIG. 2 is a block diagram for explaining the configuration of an electronic device according to one or more embodiments of the present disclosure.
[0010] FIG. 3 is a block diagram illustrating additional configurations of an electronic device and a server according to one or more embodiments of the present disclosure.
[0011] FIGS. 4, FIGS. 5 and FIG. 6 are drawings for illustrating a similarity map showing the similarity between a plurality of objects according to one or more embodiments of the present disclosure.
[0012] FIG. 7 is a diagram illustrating the layer hierarchy of an artificial intelligence model according to one or more embodiments of the present disclosure.
[0013] FIG. 8 is a drawing for explaining the gradual adjustment of learning difficulty according to one or more embodiments of the present disclosure.
[0014] FIG. 9 is a drawing for explaining the configuration of a server according to one or more embodiments of the present disclosure.
[0015] FIG. 10 is a flowchart illustrating a learning method of an electronic device according to one or more embodiments of the present disclosure.
[0016] The embodiments described herein are subject to various modifications and may have various forms; specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the scope of specific embodiments and should be understood to include various modifications, equivalents, and / or alternatives of the embodiments of the present disclosure. In relation to the description of the drawings, similar reference numerals may be used for similar components.
[0017] In describing the present disclosure, if it is determined that a detailed description of related known functions or configurations could unnecessarily obscure the essence of the present disclosure, such detailed description is omitted.
[0018] Additionally, the following embodiments may be modified in various other forms, and the scope of the technical concept of the present disclosure is not limited to the following embodiments. Rather, these embodiments are provided to make the present disclosure more faithful and complete and to fully convey the technical concept of the present disclosure to those skilled in the art.
[0019] The terms used in this disclosure are used merely to describe specific embodiments and are not intended to limit the scope of the rights. The singular expression includes the plural expression unless the context clearly indicates otherwise.
[0020] In the present disclosure, expressions such as “have,” “may have,” “include,” or “may include” indicate the presence of such features (e.g., numerical values, functions, actions, or components such as parts) and do not exclude the presence of additional features.
[0021] In the present disclosure, expressions such as “A or B,” “at least one of A or / and B,” or “one or more of A or / and B” may include all possible combinations of items listed together. For example, “A or B,” “at least one of A and B,” or “at least one of A or B” may refer to cases including (1) at least one A, (2) at least one B, or (3) both at least one A and at least one B.
[0022] Expressions such as "first," "second," "first," or "second" used in this disclosure may modify various components regardless of order and / or importance, and are used only to distinguish one component from another and do not limit said components.
[0023] When it is stated that a certain component (e.g., a first component) is "(operatively or communicatively) coupled with / to" or "connected to" another component (e.g., a second component), it should be understood that the said certain component may be directly connected to the said other component or connected through another component (e.g., a third component).
[0024] On the other hand, when it is stated that a certain component (e.g., a first component) is "directly connected" or "directly coupled" to another component (e.g., a second component), it may be understood that no other component (e.g., a third component) exists between said certain component and said other component.
[0025] As used in this disclosure, the expression “configured to” may be replaced, depending on the context, with, for example, “suitable for,” “having the capacity to,” “designed to,” “adapted to,” “made to,” or “capable of.” The term “configured to” may not necessarily mean only “specifically designed to” in hardware.
[0026] Instead, in some situations, the expression “device configured to do something” may mean that the device is “capable of doing something” in conjunction with other devices or components. For example, the phrase “processor configured (or set) to perform A, B, and C” may refer to a dedicated processor for performing those operations (e.g., an embedded processor), or a generic-purpose processor (e.g., a CPU or application processor) capable of performing those operations by executing one or more software programs stored in a memory device.
[0027] In the embodiments, a 'module' or 'part' performs at least one function or operation and may be implemented in hardware or software, or a combination of hardware and software. Additionally, a plurality of 'modules' or a plurality of 'parts' may be integrated into at least one module and implemented by at least one processor, except for the 'module' or 'part' that needs to be implemented in specific hardware.
[0028] Meanwhile, various elements and areas in the drawings are depicted schematically. Accordingly, the technical concept of the present disclosure is not limited by the relative sizes or spacing depicted in the attached drawings.
[0029] Hereinafter, embodiments according to the present disclosure are described in detail with reference to the attached drawings so that those skilled in the art can easily implement them.
[0030] FIG. 1 is a drawing for explaining the operation of an electronic device according to one or more embodiments of the present disclosure.
[0031] Although the external appearance of an electronic device (100) including a display is shown in FIG. 1, the electronic device (100) is not necessarily implemented only in a form that directly includes a display, and even if it does not directly include a display, it may be implemented in a form that outputs image data by being connected to an external display device via wired or wireless connection. Alternatively, it may be implemented as an electronic device for learning an artificial intelligence model regardless of the display.
[0032] Specifically, the electronic device (100) can be implemented in various forms such as a PC, laptop PC, smartphone, set-top box, tablet PC, TV, beam projector, server device, etc.
[0033] In the present disclosure, "display" refers to a component responsible for displaying an image, and "external display device" may refer to an external device capable of providing display functions, such as a TV, monitor, or projector. Additionally, "wired or wireless connection" may include communication methods such as HDMI, Wi-Fi, or Bluetooth.
[0034] According to FIG. 1, an electronic device (100) can display an image frame (10). The electronic device (100) can identify an object (1) included in the image frame (10) and display objects (2, 3, 4 or 2, 5, 6) corresponding to the object (1) on one side of the screen.
[0035] According to various embodiments of the present disclosure, an electronic device (100) can obtain an identification result for an object in an image using an artificial intelligence model trained using a similarity map representing the similarity between a plurality of objects. For convenience of explanation, the upper drawing of FIG. 1 shows an identification result by an artificial intelligence model that has not been trained using a similarity map, and the lower drawing shows an identification result by an artificial intelligence model that has been trained using a similarity map. A specific training method will be described in detail in the following section.
[0036] The electronic device (100) divided the entire display and displayed an image frame (10) included in the input content in the central area of the display, while displaying identification results (2, 3, 4 or 2, 5, 6) arranged vertically in the right area.
[0037] However, the screen splitting method is not necessarily limited to this and can be varied depending on various conditions such as display size, aspect ratio, user settings, and type of content.
[0038] The electronic device (100) can receive content from various content sources such as a network interface, a broadcast receiving module, a media player, and a game player, detect video content from the received content, and perform various image processing operations such as decoding and scaling on the detected video content to obtain multiple image frames. Each image frame consists of multiple pixels. The electronic device (100) can analyze the location and pixel value of the pixels within each image frame to identify an object (1) and display information about other objects having characteristics similar to those of the object (1) on one side of the screen.
[0039] The electronic device (100) can perform this object identification task using an artificial intelligence model. The artificial intelligence model may be stored in the electronic device (100) itself, or it may be stored in an external server device or other electronic device. The electronic device (100) may input an image frame containing an object into the artificial intelligence model and obtain an identification result corresponding to the object based on the output of the artificial intelligence model. If the artificial intelligence model is stored in a server device or another external electronic device, the electronic device (100) may transmit the image frame or content to the server device or the external electronic device.
[0040] The identification result of the artificial intelligence model may include information about at least one object. The object information included in the identification result may also be described as an identified object, a candidate object, a similar object, etc. The electronic device (100) may identify an object having the highest probability value among the output values of the artificial intelligence model as the correct object, and identify an object having a lower probability value as a similar object.
[0041] The operation of acquiring multiple objects based on an image will be explained in detail in Fig. 2.
[0042] For example, FIG. 1 illustrates a state in which an electronic device (100) displays an image of a person eating food. However, this is merely one example and the present disclosure is not limited thereto, and various types of images may be input.
[0043] The electronic device (100) can input such an image into an artificial intelligence model. The artificial intelligence model analyzes the input image to obtain a similar object corresponding to an object (1) included in the image, and illustrates it as in FIG. 1. Referring to FIG. 1, when comparing the objects (2, 3, 4) identified by the artificial intelligence model not trained by the similarity map with the objects (2, 5, 6) identified by the artificial intelligence model trained by the similarity map, it can be seen that the remaining similar objects, excluding the correct object (2), are different.
[0044] For example, if the object (1) in FIG. 1 is “tomato spaghetti,” the identification results of an artificial intelligence model trained without considering a similarity map include “tomato spaghetti (2),” “pad thai (3),” and “tomato soup (4).” This is the case where an artificial intelligence model trained in a Top-1 manner without considering a similarity map is used.
[0045] The Top-1 method refers to a training method that iteratively outputs the object as the correct answer object by selecting the feature vector that is closest in distance to the feature vector of the object included in the input data.
[0046] In the present disclosure, distance refers to a numerical value representing the difference between feature vectors, serving as a criterion for expressing how semantically similar two objects are. A smaller distance value is considered to indicate a higher degree of similarity between the two objects, while a larger distance value is interpreted to indicate a greater difference between them.
[0047] According to FIG. 1, an artificial intelligence model trained without considering similarity will have lower accuracy for similar objects other than the one correct object (2) most similar to the object (1). Accordingly, Pad Thai (3), tomato soup (4), etc., which are simply similar in shape or color to tomato spaghetti (1) can be identified as similar objects.
[0048] In contrast, according to the various learning methods of the present disclosure, an artificial intelligence model is trained using a similarity map in a Top-K manner.
[0049] The Top-K method refers to a method that identifies the top K objects based on the distance between each output value from the AI model and the input object, and trains the AI model by feeding back the feature vector of each of the top K objects.
[0050] The existing Top-1 method has limitations in that it focuses solely on a single correct answer, failing to adequately reflect information about similar candidates other than the correct answer. In contrast, the Top-K method considers multiple candidates together, allowing it to incorporate multiple objects semantically similar to the input data into the learning process. Furthermore, since it can conduct training by including similar objects other than the correct answer, even when the boundaries between objects are ambiguous, it enables the construction of a more flexible model.
[0051] According to various embodiments of the present disclosure, by training an artificial intelligence model using a Top-K method that considers similarity maps, various results corresponding to user intent can be obtained, and the problem of excessive reliance on a single correct answer object can be mitigated, providing the advantage of being able to respond to various prediction situations. Consequently, the flexibility and adaptability of the artificial intelligence model to real-world environments can be improved.
[0052] The similarity maps used in various embodiments of the present disclosure will be described in detail in FIGS. 4 to 6.
[0053] In the present disclosure, a similarity map may be a dataset containing similarities between objects set according to various criteria by a learner training an artificial intelligence model. For example, the learner may set a high similarity between Italian foods such as spaghetti, pizza, risotto, and lasagna by considering regional characteristics, or conversely, set a high similarity between noodle dishes such as spaghetti, noodles, and pad Thai by considering shape characteristics.
[0054] In the example of FIG. 1, if the object (1) is tomato spaghetti, the identification result of the artificial intelligence model trained using a similarity map that considers regional characteristics can be tomato spaghetti (2), pizza (5), and risotto (6). Even within the same Italian food category, objects with similar shapes or colors can be identified with a higher rank.
[0055] As another example, if the learner in FIG. 1 trains an artificial intelligence model by assigning weights to "ingredients" among the multiple criteria of the similarity map, identification results including "tomato spaghetti (2)", "tomato soup (5)", and "tomato salad (6)" may be output. However, this is merely one example, and the identification information output may be varied depending on the method of setting weights for the multiple criteria of the similarity map or the characteristics of the input data.
[0056] As described above, when an artificial intelligence model is trained using a similarity map, a different identification result (2, 5, 6) can be obtained from the previous identification result (2, 3, 4) according to the user's preference.
[0057] FIG. 2 is a block diagram for explaining the configuration of an electronic device according to one or more embodiments of the present disclosure.
[0058] According to FIG. 2, the electronic device (100) may include memory (110) and a processor (120).
[0059] The memory (110) is configured to store various data, etc., necessary for the operation of the electronic device (100). The memory (110) may be implemented in a form embedded in the electronic device (100) or in a form combined with the electronic device (100). Here, the embedded form may mean a state in which the memory (110) is fixedly installed inside the electronic device and is integrated with the device without separate external mounting.
[0060] The memory embedded in the electronic device (100) may be implemented as at least one of volatile memory (e.g., DRAM (dynamic random access memory), SRAM (static random access memory), SDRAM (synchronous dynamic random access memory)) or non-volatile memory (e.g., OTPROM (one-time programmable ROM), PROM (programmable ROM), EPROM (erasable programmable ROM), EEPROM (electrically erasable programmable ROM), mask ROM, flash ROM (e.g., NAND flash or NOR flash), hard drive, or solid state drive (SSD). Additionally, the memory that can be attached to and detached from the electronic device (100) can be implemented in the form of a memory card (e.g., CF (compact flash), SD (secure digital), Micro-SD (micro secure digital), Mini-SD (mini secure digital), xD (extreme digital), MMC (multi-media card)) or an external memory that can be connected to a USB port (e.g., USB memory).
[0061] The memory (110) can store various data to support the operation of the electronic device (100), and, for example, a similarity map indicating the similarity between multiple objects can be stored. A detailed description of the similarity map will be provided in FIGS. 4 to 6.
[0062] According to one example, the memory (110) can store an artificial intelligence model. However, it is not limited to this, and the artificial intelligence model may also be stored in a server device or other external device. A detailed explanation of this will be provided in FIG. 3.
[0063] If the artificial intelligence model is stored in a server device, the electronic device (100) can access the server device through a communication interface to train the artificial intelligence model. Alternatively, it can call or reference the trained artificial intelligence model for use. A detailed explanation of this will be provided in FIG. 3.
[0064] At least one processor (120) is configured to perform overall control operations of the electronic device (100).
[0065] At least one processor (120) may be implemented as a digital signal processor (DSP) or microprocessor that processes digital signals. However, it is not limited thereto and may be implemented as one or more of a central processing unit (CPU), a micro controller unit (MCU), a micro processing unit (MPU), a controller, an application processor (AP), a graphics processing unit (GPU), a communication processor (CP), or an ARM processor, or as other processors capable of performing the same functions.
[0066] In addition, at least one processor (120) may be implemented as a System on Chip (SoC) or Large Scale Integration (LSI) with a built-in processing algorithm, or as a Field Programmable Gate Array (FPGA).
[0067] In addition, at least one processor (120) can perform various functions by executing computer executable instructions stored in memory.
[0068] Meanwhile, although FIG. 2 illustrates that the electronic device (100) includes only one processor, it may include multiple processors (e.g., CPU + GPU, CPU + DSP) during implementation.
[0069] According to at least one embodiment of the present disclosure, at least one processor (120) can obtain a correct answer object and at least one similar object corresponding to an object included in the input data based on output data obtained through an artificial intelligence model based on input data.
[0070] The input data may include image data as described in FIG. 1. Specifically, the image data may refer to the entire image or a part of the entire image (e.g., one or more object regions). That is, the input data may be image data containing the entire scene, and in some cases, may be a partial image containing a specific object or an image separated by region.
[0071] An artificial intelligence model is a model designed to predict or classify what objects are included in image data provided as input. The artificial intelligence model can extract features from an input image, generate a feature vector for classification based on the extracted features, and then provide the feature vector to a classification layer or an output layer to determine one or more objects to which the input image may belong. A detailed explanation of the artificial intelligence model will be provided in Fig. 9.
[0072] Output data of an artificial intelligence model may refer to prediction result data generated by processing image data input into the AI model. The output data may include probability values for various candidate objects regarding objects within the input image.
[0073] An object included in the input data may refer to one or more animals or objects existing within the input image. Although the present disclosure has focused on examples regarding food, the scope of application of the present invention is not limited thereto and can be equally applied to objects including various types or individual items, such as people, animals, objects, food, buildings, and vehicles. The object may be included in the entire image or a part of the image and refers to an object that is subject to classification, recognition, or similarity evaluation by an artificial intelligence model. An object is a visual unit that serves as a standard for prediction or classification results, and in some cases, a single image may contain multiple objects.
[0074] In the present disclosure, the term "correct answer object" refers to the object with the highest probability value, i.e., similarity value, among a plurality of output objects included in the output data. That is, the correct answer object refers to the result object determined to have the highest semantic or visual similarity with the input data. However, the correct answer object is not a fixed concept of the correct answer, but can be dynamically determined according to the interpretation criteria for the input object. A detailed explanation of the similarity map will be provided in FIGS. 4 to 6.
[0075] Similar objects refer collectively to multiple output objects included in output data obtained through an artificial intelligence model corresponding to input data, or they may be a predetermined number of objects among those excluding the correct object, ordered by their similarity values. For example, when the Top-K method is applied, similar objects may consist of the remaining K-1 objects excluding the correct object from the K candidates. These similar objects may be considered as a group of candidates that are semantically similar to the correct object. A detailed explanation of the similarity map will be provided in Figures 4 through 6.
[0076] At least one processor (120) can control at least one model parameter of an artificial intelligence model to adjust the learning loss reduction for at least one similar object based on the difference between the similarity between the correct object and at least one similar object obtained through the artificial intelligence model and the similarity included in the similarity map.
[0077] The similarity output by the artificial intelligence model for a correct object and at least one similar object may refer to a similarity based on the distance between each object included in the input data and each object included in the output data, regardless of the similarity map of the present disclosure. Since the definition of distance has been explained in FIG. 1, the distance value will be explained below. The distance value may refer to a value calculated according to, for example, the Euclidean distance, cosine distance, dot-product-based distance, or other distance calculation methods between the feature vectors of the input object and the output object. The similarity output by the artificial intelligence model operates by selecting or ranking objects with high similarity by sorting or comparing these distance-based calculation results, and can determine similarity based on the mathematical distance value itself rather than semantic similarity. Therefore, the similarity output by the artificial intelligence model may refer to a similarity determined based solely on distance relationships in a fixed feature space, without user settings or semantic-based similarity adjustments.
[0078] The difference between the similarity between a correct answer object and at least one similar object obtained through an artificial intelligence model and the similarity included in a similarity map may refer to the difference from the similarity of the artificial intelligence model that appears by adjusting similarity values based on a defined similarity map to reflect the semantic similarity between multiple output objects. Accordingly, for the same input, a general distance-based method (similarity output by an artificial intelligence model) and the similarity map-based method of the present disclosure may output different similarity values or similarity rankings, which may result in differences in the selection of output objects, determination of correct answer objects, and determination of similar objects.
[0079] The processor (120) may be configured to update internal parameters (e.g., weight, bias, etc.) to minimize the value of a loss function calculated based on the difference between the prediction result predicted for input data in the artificial intelligence model and the similarity listed in the similarity map.
[0080] The above loss function can be implemented, for example, cross-entropy, mean squared error, cosine similarity-based loss, etc. The value of the loss function, that is, the loss value, is intended to numerically reflect the accuracy of the prediction result. Based on the loss value, the processor (120) calculates the gradient of each parameter for the loss function and then updates the parameter value by applying gradient descent or a variation of the algorithm (e.g., Adam, RMSProp, Adagrad, etc.). The processor (120) performs updates repeatedly within a certain range based on a predetermined learning rate. Through this process, the artificial intelligence model can gradually reduce the loss value and improve the mapping accuracy between the input and the output. By training the artificial intelligence model by incorporating feedback on multiple similar objects based on the similarity map, objects other than the correct answer can also be identified to a degree relatively similar to the correct answer. Accordingly, search results that match the user's intent can be provided more accurately.
[0081] In addition, the learning method according to various embodiments of the present disclosure may be performed in a manner that stably converges the loss value by repeatedly inputting learning data for a certain number of epochs.
[0082] Specifically, in the present disclosure, each factor in the equation regarding total loss can be changed, and a plurality of criteria and the weights of the criteria in the similarity map can be changed. Through such changes, at least one model parameter of the artificial intelligence model can be updated to adjust the learning direction of the artificial intelligence model so as to reduce the learning loss. A detailed description of the similarity map will be provided in FIGS. 4 to 6.
[0083] The formula for total loss according to the present disclosure can be constructed as follows.
[0084] "total_loss = (1-alpha)*contrastive_loss + alpha * top_k_loss + beta * hard_negative_loss"
[0085] "total_loss" is a value determined by a total of 5 factors: "alpha", "beta", "contrastive_loss", "top_k_loss", and "hard_negative_loss".
[0086] Generally, contrastive loss refers to a loss function that, when an artificial intelligence model learns the relationship between pairs of input data, induces the model to reduce the distance between feature vectors when the two input data are identical or similar, and to expand the distance beyond a predetermined threshold when they are different or dissimilar. Contrastive loss can be calculated based on various distance calculation methods, such as Euclidean distance and cosine distance, and can contribute to simultaneously optimizing the aggregation of similar data and the separation of heterogeneous data.
[0087] "Contrastive_loss," which corresponds to one factor in the above formula for total loss, follows the definition as above. That is, "contrastive_loss" does not refer to a loss value based on the Top-K method and similarity map according to the present disclosure, but may refer to a loss value based on distance.
[0088] "top_k_loss" refers to a loss value based on the Top_K method and a similarity map. It is not merely based on distance, but is a loss value that incorporates the user's intent. In other words, it is not limited to a value calculated based solely on the distance between input and output data, but rather refers to a value that can be adjusted by reflecting a similarity map defined by the user or weighting settings of multiple criteria included in the similarity map. Therefore, the loss value of the present invention can be utilized for learning optimization in a direction that aligns with the user's intent, going beyond the results of distance-based mathematical calculations.
[0089] "hard_negative_loss" is a value related to classifying multiple objects into similar and dissimilar classes based on a similarity map, and prioritizing the training of objects among the dissimilar classes that have close distances between feature values but low similarity according to the similarity map. A detailed explanation of this will be provided in Figure 7 while explaining "hard negative mining."
[0090] "alpha" and "beta" are continuous variables having values between 0 and 1, and can be used for weight adjustment or importance reflection during the learning process or similarity calculation process. "alpha" and "beta" may be set as fixed constants or may be dynamically adjusted according to the learning stage. A detailed explanation of this will be provided in Fig. 8.
[0091] Although the electronic device (100) in FIG. 2 is illustrated as including only basic components (i.e., memory, processor, etc.), the electronic device (100) may include various additional components in addition to the components described above.
[0092] For example, the electronic device (100) may include a display.
[0093] The display is a component for displaying the operating status, notification messages, UI screens, etc. of an electronic device (100). The display can be implemented in various forms such as an LCD (Liquid Crystal Display), an OLED (Organic Light Emitting Diodes) display, or a PDP (Plasma Display Panel). The display may also include a driving circuit, a backlight unit, etc., which can be implemented in forms such as an a-si TFT (amorphous silicon thin film transistor), an LTPS (low temperature poly silicon) TFT, or an OTFT (organic TFT). Meanwhile, the display can be implemented as a touch screen combined with a touch sensor, a flexible display, a three-dimensional display, etc. Alternatively, the display may be implemented with only one or more light-emitting elements.
[0094] The electronic device (100) can change the display state of the display according to various states, such as when the electronic device (100) is turned on, when it is operating normally, when there is insufficient power, or when there is an error, so that the user can intuitively understand the state of the electronic device (100).
[0095] However, as previously stated, the display is merely one of the various embodiments, and the display configuration may be omitted in other embodiments.
[0096] FIG. 3 is a block diagram illustrating additional configurations of an electronic device and a server according to one or more embodiments of the present disclosure.
[0097] In FIG. 3, the electronic device (100) may include a communication interface (130).
[0098] The communication interface (130) may include interfaces such as HDMI port, DP, RGB, DVI, USB, and Thunderbolt for receiving video / audio signals connected to external devices such as a server (200). HDMI, DP, and Thunderbolt are ports capable of transmitting video and audio signals simultaneously. The electronic device (100) can display various content by performing various processing such as demuxing, decoding, and scaling on various signals received from external devices connected through the communication interface (130).
[0099] According to various embodiments of the present disclosure, the communication interface (130) can communicate with a server (200) in which an artificial intelligence model is stored.
[0100] The communication interface (130) may include a wireless communication module or a wired communication module, and the wireless communication module may include at least one of, for example, a Wi-Fi module, a Bluetooth module, a mobile communication module (3G, LTE, 5G, etc.), a UWB (Ultra Wide Band) module or other wireless communication modules. The wired communication module may include, for example, a LAN (Local Area Network) module, an Ethernet module, a coaxial cable or a fiber optic cable.
[0101] The processor (120) can connect to an external server (200) through a communication interface (130) and transmit input data to the server (200). The input data may consist of, for example, image data, attribute information related to a specific object, etc., and may be used as input for classification, prediction, or inference processing by an artificial intelligence model stored or running on the server (200).
[0102] The processor (120) can receive output data generated by the artificial intelligence model of the server (200) based on the input data. The output data may include a prediction class corresponding to the input data, a probability score for each class, or ranking information. This output data corresponds to the result determined by the model, but does not include whether it is correct or semantic similarity.
[0103] The processor (120) analyzes the received output data and determines the prediction result having the highest similarity to the input data as the correct answer object, and determines other prediction results that are judged to have relatively high similarity as similar objects. At this time, the similarity determination can be performed by evaluating the semantic or conceptual similarity between the input data and each output result by referring to a similarity map stored in advance.
[0104] The processor (120) can generate feedback data to improve the performance of the artificial intelligence model based on the similarity information between the correct answer object and the similar object determined in this way, and the difference between the reference value defined in the similarity map.
[0105] The generated feedback data is transmitted to the server (200) through the communication interface (130), and the server (200) can improve the accuracy and similarity judgment performance of the model by analyzing the received feedback data and adjusting the parameters of the artificial intelligence model or performing additional retraining.
[0106] This structure, rather than simply transmitting results between input and output, can have the effect of enabling continuous intelligent and user-customized learning through interaction with the server (200) by performing interpretation of output results and generating user intent-based feedback in the electronic device (100).
[0107] However, according to one embodiment, if the electronic device (100) supports on-device learning or inference, the artificial intelligence model itself may be stored in memory (110).
[0108] In this case, processing of input data can be performed independently without a network connection. Specifically, the processor (120) can train an artificial intelligence model of the memory (110) in the manner described above and use the artificial intelligence model to generate prediction, classification, or recommendation results for the input data.
[0109] Specifically, at least one processor (120) can input input data into an artificial intelligence model stored in memory (110) to obtain output data corresponding to the input data. Additionally, the processor (120) can transmit feedback data to the artificial intelligence model to adjust the model parameters based on the difference value between the similarity between the correct answer object and the at least one similar object and the similarity included in the similarity map, or based on the difference value.
[0110] FIGS. 4 to 6 are drawings for explaining a similarity map that indicates the similarity between multiple objects.
[0111] A similarity map can be a map or dataset that defines the similarity between objects according to multiple different criteria written in a similarity table. The similarity between objects can be set to a value between 0 and 1. A similarity value can be defined as a value that quantitatively represents the semantic or visual similarity between two objects. If the similarity value between two objects is 0, they can be classified as dissimilar. Conversely, if the similarity value between two objects is 1, they can be classified as identical or virtually identical.
[0112] Multiple different criteria refer to criteria that can be set in various ways according to user definitions. For example, multiple criteria may include time-based classifications such as "breakfast, lunch, dinner," meal course classifications such as "appetizer, main course, dessert," cultural classifications such as "Eastern, Western," country classifications such as "Korean, Japanese, Chinese, Italian, French," food type classifications such as "noodles, bread, soup," and attribute-based classifications such as "ingredients." Based on these multiple criteria, a similarity table can be created, and a similarity map based on the similarity table can be utilized to define semantic similarity between input data and candidate data for each criterion and to derive results from various perspectives during the learning or inference process.
[0113] Figure 4 shows a similarity table in which food 1 through food 7 are listed according to the criteria of category, sub_category, and gradient. The similarity table in Figure 4 can serve as a criterion for obtaining a similarity map in the future. Here, the class column corresponds to a column for distinguishing each food item. The category column corresponds to a column for primarily distinguishing each food item, such as salad, breakfast, sandwich, meat dish, seafood dish, etc. The sub_category column corresponds to a column for classification more detailed than the category column, such as fruit-based, vegetable-based, meat-based, healthy food, and Western-style food. Finally, the gradient column corresponds to a column for listing the ingredients of each food item.
[0114] However, criteria such as those in Fig. 4 can be determined according to various criteria depending on the user's definition method and are not limited to the example in Fig. 4. Therefore, it is also possible to add a new criterion such as country between sub_category and gradient as in Fig. 5.
[0115] The user can adjust the weights according to the field that the user considers important. For example, in FIG. 4, if the user assigns a weight of 0.4 to catregory, 0.2 to sub_category, and 0.4 to gradient, a similarity map like FIG. 6 can be obtained.
[0116] According to Figure 6, the similarity value between food_2 and food_5 is set to 0.53. This indicates that food_2 and food_5 are judged to have semantic or attribute similarity above a certain level. Specifically, food_2 and food_5 have the same category (weight=0.4) as salad, and the same gradient (weight=0.4) as lettuce and tomato.
[0117] Additionally, according to Figure 6, the similarity value between food_4 and food_6 is set to 0.56. This indicates that food_4 and food_6 are judged to have semantic or attribute similarity above a certain level. Specifically, food_4 and food_6 have the same category (weight=0.4) as meat_dish and the same gradient (weight=0.4) as bacon.
[0118] If a user wishes to increase the similarity of foods by country, they can obtain the desired similarity map by setting a high weight for each country. In the case of learning methods that rely on distance based on numerical vectors, the predicted values diverge even if the foods belong to the same country if they look different; however, if an AI model is trained based on similarity maps, foods from the same country can be inferred to be close. Through this, applications such as recommending foods from the same country to users can be made.
[0119] Unlike existing methods that reduce learning loss based simply on the distance between feature vectors, the method of reducing learning loss based on similarity values defined in a similarity map is characterized by the fact that learning is performed based on similarity relationships that reflect the user's intent. Accordingly, rather than uniform results based on a single criterion during the learning process, it provides the advantage of deriving flexible and diverse results that suit the user's purpose or situation.
[0120] FIG. 7 is a diagram illustrating the layer hierarchy of an artificial intelligence model according to one or more embodiments of the present disclosure.
[0121] A layer in an artificial intelligence model refers to extracting features from input data and
[0122] It refers to a multi-layered structure designed to perform learning and inference based on this. An artificial intelligence model can be constructed by sequentially connecting multiple layers, and each layer performs the function of processing input data in a specific manner and transmitting it to the next layer. Layers can generally be classified into an input layer, a hidden layer, and an output layer, and the hidden layer can be further composed of various types of sub-layers.
[0123] The input layer receives raw data provided from the outside (e.g., image data) into the artificial intelligence model, converts it into numerical representations (vectors, matrices, etc.), and passes it to subsequent layers.
[0124] Hidden layers perform the role of progressively transforming and abstracting input data and may include, for example, convolutional layers that perform convolution operations, activation layers that apply nonlinear activation functions, pooling layers for dimensionality reduction or enhancement, and fully connected layers that reconstruct or combine features. Each hidden layer extracts increasingly complex and abstract features from the input data; the initial layers extract simple patterns (e.g., edges, color changes, etc.), while the deeper layers are configured to recognize complex and high-dimensional meanings (e.g., object shapes, category concepts, etc.).
[0125] The output layer performs the role of outputting final results suitable for specific purposes, such as classification, prediction, or regression, based on features passed from the final hidden layer. The output layer can typically include a softmax function, a sigmoid function, or other activation functions that support probabilistic interpretation.
[0126] An artificial intelligence model according to the present disclosure can be designed to effectively extract complex features of input data based on such a layer hierarchy and to perform learning loss reduction based on a similarity map.
[0127] At least one processor (120) according to the present disclosure can input input data into an artificial intelligence model and obtain feature values of each of a plurality of objects corresponding to the input data in the final classification layer (layer n in FIG. 7) of the artificial intelligence model.
[0128] Additionally, the processor (120) may input input data into an artificial intelligence model and obtain feature values for each of the multiple objects included in the input data in an intermediate classification layer (layer k in FIG. 7) of the artificial intelligence model.
[0129] The feature value of each of multiple objects may refer to a feature-based distance value between multiple objects. For example, since curry and stew may have similar characteristics in terms of ingredient composition, cooking method, and texture, the distance between the two can be set to be relatively close. On the other hand, since curry and salad have distinct differences in cooking method, ingredients used, and texture, the distance between the two can be set to be relatively far.
[0130] At least one processor (120) can identify the similarity between multiple objects from a similarity map and classify them into similar classes and dissimilar classes based on the similarity.
[0131] The processor (120) can classify each of multiple objects into a similar class or a dissimilar class based on a specific similarity map defined and weighted by the user, regardless of feature-based distance values. For example, if a similarity map is obtained that weights "food with broth," curry and stew may have high similarity, and curry and salad may have low similarity. In this case, curry and stew are classified as a similar class, and curry and salad are classified as a dissimilar class.
[0132] At least one processor (120) can adjust the at least one model parameter to reduce the distance of feature values of objects included in the similar class and increase the distance of feature values of objects included in the dissimilar class.
[0133] Adjusting model parameters to reduce distance can be understood as part of an optimization process aimed at decreasing the learning loss value. Specifically, this refers to the process of updating the internal parameters (e.g., weights, biases, etc.) of an AI model in a direction that reduces the distance between feature vectors of ground truth objects or similar objects. Learning loss is a loss value calculated based on the distance or similarity between two feature vectors; it has the characteristic that the loss value decreases as the distance is closer and increases as the distance increases. Therefore, reducing the distance between ground truth objects and similar objects by adjusting model parameters naturally leads to a decrease in learning loss as well. This distance-based loss reduction optimization trains the model to more accurately reflect the semantic relationships between ground truth objects and similar objects, rapidly reducing large loss values early in training and gradually improving model performance by adjusting even minute distance differences later in training.
[0134] For example, the processor (120) can adjust model parameters to reduce the distance between feature values of similar classes such as curry and stew, and can adjust model parameters to increase the distance between feature values of dissimilar classes such as curry and salad. Through these adjustments, the electronic device (100) can derive recommendation results for objects that have high semantic similarity to the input data, such as curry and stew, while objects that have low similarity to the input data, such as salad, may be excluded from the recommendation results or not derived as recommendation results.
[0135] At least one processor (120) can prioritize training the artificial intelligence model on objects included in a dissimilar class that have a short distance of feature values and low similarity recorded in a similarity map. This is referred to as "hard negative mining" in this disclosure. "Hard negative mining" is a learning method related to "hard_negative_loss" described in FIG. 2 of this disclosure.
[0136] Among objects in the dissimilar class, those with close feature distances and low similarity recorded in the similarity map are difficult to learn. This is because the AI model learns to artificially increase the distance between two objects that it inherently recognizes as similar.
[0137] For example, although the distance between feature values of curry and stew is close, if the user assigns weight to the criterion of "food that can be eaten with curry," the similarity recorded in the similarity map of curry and stew may be low. In this case, the processor (120) can prioritize training curry and stew.
[0138] When configured to prioritize the training of high-difficulty objects in the early stages, the following technical advantages can be provided. First, by having the model perform sufficient training on high-difficulty objects first, it can form clearer boundaries between objects within the feature space. This has the effect of inducing the model to establish more precise judgment criteria for cases prone to errors during the early stages of training. Additionally, training high-difficulty objects first allows subsequent training on relatively lower-difficulty objects to proceed more quickly and stably, contributing to improved overall convergence speed and enhanced generalization performance. Consequently, adjusting the order of object training based on difficulty is advantageous in terms of training efficiency, boundary clarity, and prediction stability, and can help implement more reliable AI models in various environments.
[0139] FIG. 8 is a drawing for explaining the gradual adjustment of learning difficulty according to one or more embodiments of the present disclosure.
[0140] An epoch refers to a unit in the training process of an artificial intelligence model where the entire training dataset is input into the model to perform training. The epoch count can be used as the number of training iterations; while increasing the number of epochs allows the model to learn from more data and improve performance, an excessively large number of epochs can lead to overfitting. Therefore, setting an appropriate number of epochs plays an important role in improving the model's generalization performance.
[0141] FIG. 8 illustrates an example in which a total of 11 epochs are performed. However, this is merely an example to aid in understanding the present disclosure, and in actual implementation, the number of epochs may be arbitrarily adjusted depending on the learning purpose, data characteristics, or model structure.
[0142] As "alpha", "contrastive_loss", and "top_k_loss" have already been explained, a detailed explanation is omitted.
[0143] According to Fig. 8, in the equation "total_loss = (1-alpha)*contrastive_loss + alpha * top_k_loss + beta * hard_negative_loss", when the epoch is 1 (alpha=0), "top_k_loss" becomes 0. That is, when the epoch is 1, training proceeds only in the "Top-1 method". When the epoch is 11 (alpha=1), "top_k_loss" becomes 1. That is, when the epoch is 11, training proceeds only in the "Top-K method".
[0144] The learning method of an artificial intelligence model according to the present disclosure may be configured to learn simple relationships (e.g., Top-1) in the initial stage of learning and gradually learn complex relationships (e.g., Top-K). That is, learning may be performed initially using only the Top-1 method, and then learning may be performed by gradually expanding the Top-K method. This learning method can mitigate instability in the initial stage of learning and stably guide the entire learning process by gradually adjusting the learning difficulty.
[0145] FIG. 9 is a drawing for explaining the configuration of a server according to one or more embodiments of the present disclosure.
[0146] The server (200) may include memory (210), a processor (220), and an AI model (300) stored in memory (210).
[0147] The memory (210) is a configuration that stores various programs, data, and information necessary for the operation of the server (200). Specifically, the memory (210) can store an artificial intelligence model, input data, output data, training dataset, artificial intelligence model parameters, and parameter update information.
[0148] The memory (210) of FIG. 9 is shown as being built into a server, but is not limited thereto and may be provided in a form that operates by being connected to an external storage device.
[0149] The embodiments according to the present disclosure are not limited to the type, size, or storage method of information stored in the memory (210) and can be changed according to various hardware and software configurations.
[0150] The processor (220) is configured to control and perform various functions of the server (200) by executing a program or data stored in memory (210). Specifically, the processor (220) can control the artificial intelligence model to process input data and generate output data. Additionally, it can control the artificial intelligence model to analyze the relationship between input data and output data, control the artificial intelligence model to update parameters based on received feedback data, or control the artificial intelligence model to perform retraining if necessary. The processor (220) can also perform a communication control function to receive input data from an external electronic device (e.g., electronic device (100)) and transmit the processing results of the artificial intelligence model back.
[0151] Additionally, the processor (220) can control the flow of data between multiple components within the server (200) and perform management and updates of data stored in memory (210).
[0152] The embodiments according to the present disclosure are not limited to the number, arrangement structure, or operation method of the processor (220) and can be changed according to various hardware and software structures.
[0153] The artificial intelligence model (300) stored in memory (210) can be used to perform operations such as classification, prediction, and recommendation by processing input data on the server (200) side, or can be utilized to adjust the model's parameters or perform retraining by receiving feedback data.
[0154] The artificial intelligence model (300) stored in memory can be used to process input data and perform various operations on the server (200) side. Specifically, the artificial intelligence model (300) can perform operations such as classification, prediction, anomaly detection, or clustering based on the input data. A classification operation refers to the task of identifying the input data into one of a plurality of predefined classes, and a prediction operation refers to the task of calculating a future state or a continuous value based on the input data. A recommendation operation refers to the task of suggesting one or more candidates based on similarity or preference with respect to the input data, and an anomaly detection operation may include the task of determining whether the input data deviates from a normal distribution.
[0155] The artificial intelligence model (300) can perform learning to update the model's parameters or partially update the model structure based on feedback data received from the server (200). If necessary, it may also perform retraining of the entire model or partial fine-tuning. The description of the parameter update of the artificial intelligence model (300) is detailed in FIG. 2, so the description is omitted.
[0156] The artificial intelligence model (300) may be a pre-trained model or an adaptive model that is continuously updated according to the user's input patterns or environmental changes. Additionally, it may have a single model structure or be a modular form composed of multiple sub-models.
[0157] The embodiments according to the present disclosure are not limited to the configuration, learning method, application target, or operation method of the artificial intelligence model (300), and may include various types of deep learning models (e.g., CNN, RNN, Transformer, GNN, etc.) or machine learning models (e.g., SVM, Random Forest, etc.).
[0158] As described in the preceding drawings, if a similarity map is stored in an electronic device (100) and a processor (120) of the electronic device (100) acquires and transmits learning feedback data based on the similarity map, the processor (220) can train an artificial intelligence model (300) using the received feedback data.
[0159] Conversely, if both the similarity map and the artificial intelligence model (300) are stored in the memory (210) of the server (200), the aforementioned learning method may be performed by the processor (220) of the server (200) itself. Specific examples of similarity maps and learning methods have been specifically explained in the above section, so a redundant explanation is omitted.
[0160] FIG. 10 is a flowchart illustrating a learning method of an electronic device according to one or more embodiments of the present disclosure.
[0161] The learning method of the electronic device illustrated in FIG. 10 may be performed by an electronic device (100) according to various embodiments described in FIG. 1 to 9 above, but is not necessarily limited thereto, and may also be performed by an electronic device having a structure in which some components are omitted, added, or modified. In addition, it may be performed by a set-top box or other control device connected to the electronic device (100) illustrated in FIG. 1 to 9.
[0162] The following description is based on the case where it is performed in an electronic device (100) having the configuration of Fig. 2.
[0163] The electronic device (100) can obtain output data output from an artificial intelligence model based on input data containing an object (S1010).
[0164] According to one or more embodiments, the input data and output data may include image data. The input image data may include image data related to an object corresponding to a part of the input image data. The output image data may include image data related to an object corresponding to a part of the output image data.
[0165] The electronic device (100) can obtain a correct answer object corresponding to the object and at least one similar object based on output data (S1020).
[0166] The description of the correct answer object and at least one similar object has been explained in detail previously, so it is omitted.
[0167] The electronic device (100) can identify the similarity between a correct object and at least one similar object based on a similarity map corresponding to the similarity between a plurality of objects (S1030).
[0168] The explanation regarding similarity map similarity identification has been explained in detail previously, so it will be omitted.
[0169] The electronic device (100) can control at least one model parameter of the artificial intelligence model to be adjusted so as to reduce the learning loss for at least one similar object based on the difference between the similarity between the correct object and at least one similar object obtained through the artificial intelligence model and the similarity identified from the similarity map (S1040).
[0170] The explanation of the learning loss and the model parameters of the artificial intelligence model has been explained in detail previously, so it will be omitted.
[0171] Although various embodiments of the present disclosure have been described in detail for each embodiment, each embodiment is not required to be implemented alone and may be implemented in part or wholly with other embodiments.
[0172] Meanwhile, methods according to at least some of the various embodiments of the present disclosure described above may be implemented in the form of an application installable on an existing electronic device. Additionally, methods according to at least some of the various embodiments of the present disclosure described above may be implemented solely through a software upgrade or a hardware upgrade of an existing electronic device.
[0173] Meanwhile, according to one embodiment of the present disclosure, the various embodiments described above may be implemented as software containing instructions stored on a machine-readable storage media. The machine may include an electronic device (e.g., electronic device (A)) according to the disclosed embodiments, which is a device capable of calling instructions stored from the storage media and operating according to the called instructions. When instructions are executed by a processor, the processor may perform a function corresponding to the instructions directly or by using other components under the control of the processor. Instructions may include code generated or executed by a compiler or an interpreter. The machine-readable storage media may be provided in the form of a non-transitory storage media. Here, "non-transitory storage media" simply means that it is a tangible device and does not contain a signal (e.g., electromagnetic waves), and this term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily in the storage media. For example, a 'non-transient storage medium' may include a buffer in which data is temporarily stored. According to one embodiment, the method according to the various embodiments disclosed herein may be provided by being included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store (e.g., Play Store) or directly between two user devices (e.g., smartphones).In the case of online distribution, at least a portion of a computer program product (e.g., a downloadable app) may be temporarily stored or temporarily created on a device-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.
[0174] Although preferred embodiments of the present disclosure have been illustrated and described above, the present disclosure is not limited to the specific embodiments described above. It is understood that various modifications can be made by those skilled in the art without departing from the essence of the present disclosure as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present disclosure.
Claims
1. In an electronic device, Memory storing a map corresponding to the similarity between multiple objects; and Includes at least one processor; and The above-mentioned at least one processor is, Based on output data obtained through an artificial intelligence model in relation to input data containing an object, a correct object and at least one other object having a lower similarity to the object than the correct object are obtained, and An electronic device that controls at least one model parameter of an artificial intelligence model to adjust the learning loss associated with the at least one other object based on the difference between the similarity between the correct answer object and the at least one other object obtained through the artificial intelligence model and the similarity included in a map containing the object.
2. In Paragraph 1, A communication interface for communicating with a server where the above artificial intelligence model is stored; including, The above-mentioned at least one processor is, The input data is transmitted to the server through the communication interface, and The output data obtained through the above artificial intelligence model is received from the above server, and An electronic device that transmits feedback data to a server through a communication interface, corresponding to the adjustment of at least one model parameter based on the difference value between the similarity between the correct answer object and at least one other object and the similarity included in the map, or the difference value.
3. In Paragraph 1, The map above is It is one of multiple maps that define the similarity between each object according to multiple different criteria, and The above-mentioned at least one processor is, An electronic device that obtains the similarity of at least one other object by applying different weights to each of the plurality of maps according to user preference, and controls the at least one model parameter to be adjusted based on the difference from the similarity of at least one other object provided by the artificial intelligence model.
4. In Paragraph 1, The above-mentioned at least one processor is, Identify the object with the highest output value among the output data of the above artificial intelligence model as the above correct answer object, and An electronic device that identifies a preset number of objects included in a map corresponding to the similarity between multiple objects among the remaining objects excluding the above correct answer object as at least one other object.
5. In Paragraph 1, The above-mentioned at least one processor is, Based on the above input data, the feature values of each of the plurality of objects included in the input data are obtained from the AI model in the final classification layer of the AI model, and Based on the similarity between the plurality of objects obtained from the map above, they are classified into similar classes and dissimilar classes, and An electronic device that controls the adjustment of at least one model parameter such that the distance between feature values of objects included in the similar class decreases and the distance between feature values of objects included in the dissimilar class increases.
6. In Paragraph 5, The above-mentioned at least one processor is, An electronic device that controls the artificial intelligence model to learn preferentially for objects included in the above dissimilar class, wherein the distance between the feature values is close and the similarity is low for objects included in the above map.
7. In Paragraph 1, The above-mentioned at least one processor is, Based on the above input data, the feature values of each of the plurality of objects included in the input data are obtained from the artificial intelligence model in the intermediate layer among the plurality of layers included in the artificial intelligence model, and Based on the similarity between the plurality of objects obtained from the map above, they are classified into similar classes and dissimilar classes, and An electronic device that controls the adjustment of at least one model parameter such that the distance between feature values of a plurality of objects included in the similar class decreases and the distance between feature values of objects included in the dissimilar class increases.
8. In a method for training an electronic device for training an artificial intelligence model, A step of obtaining output data from the artificial intelligence model in relation to input data containing an object; A step of obtaining the correct answer object and at least one other object having a lower similarity to the object than the correct answer object, based on the above output data; A step of identifying the similarity between the correct answer object obtained through the artificial intelligence model and the at least one other object based on a map corresponding to the similarity between a plurality of objects; and A learning method comprising: a step of controlling at least one model parameter of the artificial intelligence model to adjust the learning loss associated with the at least one other object based on the difference between the similarity between the correct answer object and the at least one other object obtained through the artificial intelligence model and the similarity included in the map, so as to reduce the learning loss associated with the at least one other object.
9. In Paragraph 8, The step of obtaining output data from the above artificial intelligence model is, A step of transmitting the input data to a server where the artificial intelligence model is stored; The step of receiving the output data provided by the artificial intelligence model from the server; is included, The step of controlling to adjust at least one model parameter is, A learning method comprising the step of transmitting to the server feedback data corresponding to the adjustment of the at least one model parameter based on the difference value between the similarity between the correct answer object and the at least one other object and the similarity included in the map, or the difference value.
10. In Paragraph 8, The map above is It includes multiple maps that define the similarity between each object according to multiple different criteria, and The control step for adjusting at least one model parameter is as follows: A learning method comprising: a step of obtaining the similarity of at least one other object by applying different weights to each of the plurality of maps according to user preference, and controlling the at least one model parameter to be adjusted based on the difference from the similarity of at least one other object provided by the artificial intelligence model.
11. In Paragraph 8, The step of acquiring the above-mentioned correct answer object and at least one other object is, Identify the object with the highest output value among the output data of the above artificial intelligence model as the above correct answer object, and A learning method comprising: a step of identifying a predetermined number of objects included in a map corresponding to the similarity between multiple objects among the remaining objects excluding the above correct answer object as at least one other object.
12. In Paragraph 8, The step of obtaining output data from the above artificial intelligence model is, The method includes the step of obtaining feature values of each of a plurality of objects included in the input data from the artificial intelligence model based on the input data, in the final classification layer of the artificial intelligence model. The step of identifying similarity between at least one other object is, The step of classifying into similar classes and dissimilar classes based on the similarity between the plurality of objects obtained from the map; The step of controlling to adjust at least one model parameter is, A learning method comprising: a step of controlling the adjustment of at least one model parameter such that the distance between feature values of objects included in the similar class decreases and the distance between feature values of objects included in the dissimilar class increases.
13. In Paragraph 12, The step of controlling to adjust at least one model parameter is, A learning method comprising: a step of controlling the artificial intelligence model to prioritize learning of objects included in the dissimilar class that have a short distance between feature values and low similarity included in the map.
14. In Paragraph 8, The step of obtaining output data from an artificial intelligence model is, The method includes the step of obtaining feature values of each of a plurality of objects included in the input data from an intermediate layer among a plurality of layers included in the artificial intelligence model based on the input data; The step of identifying similarity between at least one other object is, The step of classifying into similar classes and dissimilar classes based on the similarity between the plurality of objects obtained from the map; The step of controlling to adjust at least one model parameter is, A learning method comprising: a step of controlling the adjustment of at least one model parameter such that the distance between feature values of objects included in the similar class decreases and the distance between feature values of objects included in the dissimilar class increases.
15. A non-transient computer-readable recording medium storing a program for executing a learning method of an electronic device for learning an artificial intelligence model, The above learning method is, A step of obtaining output data from the artificial intelligence model in relation to input data containing an object; A step of obtaining the correct answer object and at least one other object having a lower similarity to the object than the correct answer object, based on the above output data; A step of identifying the similarity between the correct answer object obtained through the artificial intelligence model and the at least one other object based on a map corresponding to the similarity between multiple objects; A non-transient computer-readable recording medium comprising: a step of controlling to adjust at least one model parameter of the artificial intelligence model so as to reduce the learning loss associated with the at least one other object, based on the difference between the similarity between the correct answer object and the at least one other object obtained through the artificial intelligence model and the similarity included in the map.
