Method and electronic device for improving accuracy of artificial intelligence

Abstract

According to the present disclosure, an artificial intelligence (AI) accuracy improvement method of an electronic device is disclosed, the method comprising the steps of: designing AI based on a Monte-Carlo simulation algorithm; setting a region of interest for executing a local Monte Carlo simulation within a vulnerable region in which the AI is determined to be vulnerable from among application regions of the AI; executing a local Monte Carlo simulation for the region of interest on the basis of the AI; and identifying information about the improved accuracy of the AI on the basis of the execution of the local Monte Carlo simulation for the region of interest, wherein the improved accuracy is determined on the basis that deep learning data acquired through the execution of the local Monte Carlo simulation for the region of interest is set to be additionally reflected in deep learning data pre-acquired through the execution of the Monte Carlo simulation for the application region.

Description

A METHOD FOR IMPROVING ACCURACY OF ARTIFICIAL INTELLIGENCE AND ELECTRONIC APPARATUS THEREOF

[0001] The present invention relates to a method and device for improving the accuracy of artificial intelligence, and more particularly, to a method and electronic device for improving the accuracy of artificial intelligence by performing local simulations on vulnerable areas of artificial intelligence.

[0002] When designing AI (Artificial Intelligence) systems, when simulations of numerous scenarios at each stage are impossible, Monte Carlo algorithms, a random sampling method, are often used to approximate the results. Thus, Monte Carlo algorithms are widely used in AI development to find optimal solutions among big data.

[0003] For example, in the case of Go AI services, Monte-Carlo Tree Search (MCTS) determines the optimal next move. It assumes both the player and the opponent have identical policy networks, runs multiple simulations, and selects the move with the highest frequency. This method is known to yield near-optimal results if sufficient simulations are performed on high-performance equipment, without having to search the entire tree. In fact, Google supports parallel processing and allocates numerous CPUs and GPUs to ensure sufficient simulations.

[0004] However, the MCTS algorithm is inefficient because it relies on a random sampling method and requires a significant amount of trial-and-error sampling. Therefore, current MCTS-based AI services are difficult to implement without hardware specifications such as high-performance GPUs. This means that faithful implementation of AI services requires significant costs and time.

[0005] For example, in the case of the Go AI, in the fourth game, the only one won by Lee Sedol out of five matches between him and AlphaGo, Lee Sedol, on his 78th move, when the game was still relatively early in the game, initiated a close-quarters battle in the center, where the maximum number of escape paths was four, with stones placed tightly together with no gaps. The number of possible central escape paths in the battle was extremely large, consisting mainly of multiples of four, such as "4×4×4×...". This greatly weakened AlphaGo's MCTS-based AI, which determines the breadth of its move options and the depth of its next move, allowing Lee Sedol to win. To put it more simply, Lee Sedol's human eyes could easily see the center of the board, making it easy to predict the possible outcomes in the battle. However, AlphaGo's MCTS-based vision could interpret the center as a deep, black hole, making it difficult to see. Ultimately, in the case of Baduk AI, the central area is an area that is not easily visible due to the large number of cases in the early stages, so it can be said that it is vulnerable to methods that induce a large number of cases in the early stages.

[0006] To address the fundamental problem of AI handling "a very large number of possible outcomes," as described above, traditionally, AlphaGo and other Go AIs have adopted neural network structures, including value networks and policy networks. However, even with these neural network concepts, Go AIs still struggle with accuracy, such as inability to accurately predict winning odds for large-scale battles in vulnerable areas like the central region.

[0007] According to the method of the present invention, an electronic device can improve the accuracy of artificial intelligence by performing local simulations on vulnerable areas of artificial intelligence.

[0008] The technical problems to be achieved in the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention belongs from the description below.

[0009] Various embodiments may provide methods and electronic devices for improving artificial intelligence (AI) accuracy.

[0010] A method for improving the accuracy of artificial intelligence (AI) of an electronic device according to various embodiments comprises: designing an AI based on a Monte Carlo simulation algorithm; setting a region of interest for executing a local Monte Carlo simulation within a vulnerable region of the AI ​​determined to be vulnerable among application regions of the AI; executing a local Monte Carlo simulation for the region of interest based on the AI; and confirming information on improved accuracy of the AI ​​based on the execution of the local Monte Carlo simulation for the region of interest, wherein the improved accuracy can be determined based on a point that deep learning data acquired through execution of the local Monte Carlo simulation for the region of interest is additionally reflected in deep learning data already acquired through execution of the Monte Carlo simulation for the application region.

[0011] In an exemplary embodiment, the AI ​​may be configured to apply a Latin Hypercube algorithm instead of the Monte Carlo simulation algorithm based on conditions related to uniformity and efficiency.

[0012] In an exemplary embodiment, when the AI ​​corresponds to a Go AI, an M*M line region of interest is set within a central portion of the entire Go board area of ​​N*N lines where the AI ​​is determined to be vulnerable, and the improved accuracy is determined based on the point that deep learning data acquired through execution of a local Monte Carlo simulation for the M*M line region of interest is additionally reflected in deep learning data acquired through execution of a Monte Carlo simulation for the entire Go board area of ​​N*N lines, and the M may be smaller than the N.

[0013] In an exemplary embodiment, when the AI ​​corresponds to a medical diagnosis AI, a medical diagnosis area of ​​interest is set in which the medical error judgment rate by the AI ​​is determined to be high among the medical diagnosis areas, and the improved accuracy can be determined based on the fact that deep learning data acquired through execution of a local Monte Carlo simulation for the medical diagnosis area of ​​interest is additionally reflected in deep learning data already acquired through execution of a Monte Carlo simulation for the medical diagnosis area.

[0014] In an exemplary embodiment, if it is determined that the AI ​​is in an application error state for a specific area among the application areas, the AI ​​may be set to exclude the application result of the AI ​​for the specific area based on a processing algorithm preset in the AI ​​and obtain only the application result of the AI ​​for the remaining areas among the application areas excluding the specific area.

[0015] In an exemplary embodiment, the method for improving AI accuracy may further include: setting a plurality of regions of interest having different sizes within the vulnerable region; for each region of interest included in the plurality of regions of interest, a step of confirming each improved accuracy determined for the AI ​​based on execution of a local Monte Carlo simulation for each region of interest; and a step of setting information including the size of each region of interest and each improved accuracy.

[0016] In an exemplary embodiment, the method for improving AI accuracy may further include: a step of checking the degree of each computational load determined when executing a local Monte Carlo simulation for each region of interest included in the plurality of regions of interest; and a step of determining a region of interest of an optimal size for the performance of the AI ​​among the plurality of regions of interest based on the respective improved accuracy and respective degrees of computational load.

[0017] An electronic device for performing a method for improving the accuracy of artificial intelligence (AI) according to various embodiments, comprising: a processor; and one or more memories storing one or more instructions, wherein the one or more instructions, when executed, control the processor to perform: designing AI based on a Monte Carlo simulation algorithm; setting a region of interest for executing a local Monte Carlo simulation within a vulnerable region of the AI ​​determined to be vulnerable among the application regions of the AI; executing a local Monte Carlo simulation for the region of interest based on the AI; and checking information on improved accuracy of the AI ​​based on the execution of the local Monte Carlo simulation for the region of interest, wherein the improved accuracy can be determined based on a point that deep learning data acquired through execution of the local Monte Carlo simulation for the region of interest is additionally reflected in deep learning data already acquired through execution of the Monte Carlo simulation for the application region.

[0018] In an exemplary embodiment, the AI ​​may be configured to apply a Latin Hypercube algorithm instead of the Monte Carlo simulation algorithm based on conditions related to uniformity and efficiency.

[0019] In an exemplary embodiment, when the AI ​​corresponds to a Go AI, an M*M line region of interest is set within a central portion of the entire Go board area of ​​N*N lines where the AI ​​is determined to be vulnerable, and the improved accuracy is determined based on the point that deep learning data acquired through execution of a local Monte Carlo simulation for the M*M line region of interest is additionally reflected in deep learning data acquired through execution of a Monte Carlo simulation for the entire Go board area of ​​N*N lines, and the M may be smaller than the N.

[0020] In an exemplary embodiment, when the AI ​​corresponds to a medical diagnosis AI, a medical diagnosis area of ​​interest is set in which the medical error judgment rate by the AI ​​is determined to be high among the medical diagnosis areas, and the improved accuracy can be determined based on the fact that deep learning data acquired through execution of a local Monte Carlo simulation for the medical diagnosis area of ​​interest is additionally reflected in deep learning data already acquired through execution of a Monte Carlo simulation for the medical diagnosis area.

[0021] In an exemplary embodiment, if it is determined that the AI ​​is in an application error state for a specific area among the application areas, the AI ​​may be set to exclude the application result of the AI ​​for the specific area based on a processing algorithm preset in the AI ​​and obtain only the application result of the AI ​​for the remaining areas among the application areas excluding the specific area.

[0022] In an exemplary embodiment, the method for improving AI accuracy may further include: setting a plurality of regions of interest having different sizes within the vulnerable region; for each region of interest included in the plurality of regions of interest, a step of confirming each improved accuracy determined for the AI ​​based on execution of a local Monte Carlo simulation for each region of interest; and a step of setting information including the size of each region of interest and each improved accuracy.

[0023] In an exemplary embodiment, the method for improving AI accuracy may further include: a step of checking the degree of each computational load determined when executing a local Monte Carlo simulation for each region of interest included in the plurality of regions of interest; and a step of determining a region of interest of an optimal size for the performance of the AI ​​among the plurality of regions of interest based on the respective improved accuracy and respective degrees of computational load.

[0024] The various embodiments of the present disclosure described above are only some of the preferred embodiments of the present disclosure, and various embodiments reflecting the technical features of the various embodiments of the present disclosure can be derived and understood by a person having ordinary skill in the art based on the detailed description to be described below.

[0025] The present invention has a technical effect in that it can improve the accuracy of artificial intelligence by allowing an electronic device to set a region of interest within a vulnerable area of ​​artificial intelligence and to reflect deep learning data obtained through local simulation execution on the region of interest in the operation of artificial intelligence.

[0026] The effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by a person having ordinary skill in the art to which the present invention belongs from the description below.

[0027] FIG. 1 is a diagram illustrating the configuration of an electronic device that performs the artificial intelligence accuracy improvement method of the present disclosure.

[0028] Hereinafter, various embodiments of the present disclosure will be described in detail with reference to the attached drawings so that those skilled in the art can easily implement the present disclosure. The detailed description provided below is intended to explain exemplary embodiments of the various embodiments and is not intended to be limited to any specific embodiment.

[0029] The terms used in the examples have been selected from widely used, current terms, taking into account the functions of the present disclosure. However, these terms may vary depending on the intentions of those skilled in the art, precedents, the emergence of new technologies, etc. Furthermore, in certain cases, terms may be arbitrarily selected by the applicant, in which case their meanings will be described in detail in the relevant description. Therefore, the terms used in this disclosure should not be defined simply as names, but rather based on the meanings of the terms and the overall content of the present disclosure.

[0030] The following embodiments combine components and features of various embodiments in a predetermined form. Each component or feature may be considered optional unless explicitly stated otherwise. Each component or feature may be implemented without being combined with other components or features. Furthermore, various embodiments may be formed by combining some components and features. The order of operations described in various embodiments may be changed. Some components or features of one embodiment may be included in another embodiment or may be replaced with corresponding components or features of another embodiment.

[0031] In the description of the drawings, procedures or steps that may obscure the gist of various embodiments are not described, and procedures or steps that can be understood by a person with ordinary knowledge in the relevant technical field are also not described.

[0032] Throughout the specification, when a part is said to "comprising" or "including" a component, this does not mean excluding other components, but rather that it may include other components, unless the contrary is specifically stated. The singular form of a noun corresponding to an item in this document may be used in both the singular and plural sense, unless the specification indicates otherwise or the context clearly contradicts it. In this document, phrases such as "A or B," "at least one of A and B," "at least one of A or B," "A, B, or C," "at least one of A, B, and C," and "at least one of A, B, or C" each include any one of the items listed in that phrase, or all possible combinations thereof. Terms such as "first," "second," or "first" or "second" may be used merely to distinguish one component from another, and do not limit the components in any other respect (e.g., importance or order).

[0033] Each component (e.g., a module or a program) described in this document may include one or more entities. According to various embodiments, one or more components or operations of the components may be omitted, or one or more other components or operations may be added. Additionally, multiple components (e.g., modules or programs) may be integrated into a single component. In this case, the integrated component may perform one or more functions of each of the multiple components identically or similarly to those performed by the corresponding component among the multiple components prior to integration. According to various embodiments, the operations performed by a module, program, or other component may be executed sequentially, in parallel, or repeatedly, or one or more of the operations may be executed in a different order, omitted, or one or more other operations may be added.

[0034] The term "module" or "part" used in this document means a unit that processes at least one function or operation, and may include a unit implemented by hardware, software, firmware, or a combination thereof.

[0035] Various embodiments of this document may be implemented as software (e.g., a program or application) including one or more instructions stored in a machine-readable storage medium (e.g., memory). For example, a processor of the device may call at least one instruction among the one or more instructions stored from the storage medium and execute it. This enables the device to operate to perform at least one function according to the at least one instruction called. The one or more instructions may include code generated by a compiler or code executable by an interpreter. The machine-readable storage medium may be provided in the form of a non-transitory storage medium. Here, "non-transitory" only means that the storage medium is a tangible device and does not contain signals, and does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily.

[0036] Additionally, specific terms used in various embodiments are provided to aid understanding of the various embodiments, and the use of such specific terms may be changed in other forms without departing from the technical spirit of the various embodiments.

[0037] FIG. 1 is a diagram illustrating the configuration of an electronic device that performs the artificial intelligence accuracy improvement method of the present disclosure.

[0038] Referring to FIG. 1, an electronic device (100) may include, according to one embodiment, a processor (110) and a memory (120). The electronic device (100) illustrated in FIG. 1 only illustrates components related to the present embodiment, and it will be understood by those skilled in the art related to the present embodiment that other general components may be included in addition to the components illustrated in FIG. 1.

[0039] For example, the electronic device (100) may include a communication device including one or more transceivers, an input unit, and an output unit. The communication unit is a device for performing wired / wireless communication and may communicate with an external electronic device. The external electronic device may be a terminal or a server. In addition, communication technologies used by the communication unit may include GSM (Global System for Mobile communication), CDMA (Code Division Multi Access), LTE (Long Term Evolution), 5G, WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Bluetooth, RFID (Radio Frequency Identification), Infrared Data Association (IrDA), ZigBee, NFC (Near Field Communication), etc. The input unit may be, for example, a traditional keypad or keyboard, a mouse, a microphone for receiving voice signals, a camera, and various other input means for detecting or receiving various types of user input. The output unit may be, for example, a display that outputs images, a speaker that outputs sounds, a haptic device that generates vibrations, and various other forms of output means.

[0040] Additionally, at least some of the components within the electronic device (100) may be implemented as an integrated unit or implemented as a single or multiple units. At least some of the components within the electronic device (100) may be connected to each other via a bus, a general purpose input / output (GPIO), a serial peripheral interface (SPI), or a mobile industry processor interface (MIPI), and may exchange data and / or signals.

[0041] The processor (110) is a component that can perform calculations or data processing related to control and / or communication of each component of the electronic device (100). For example, the processor (110) can control the electronic device (100) as a whole by executing programs stored in the memory (120) within the electronic device (100). The processor (110) may be implemented as a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), etc., provided within the electronic device (100), but is not limited thereto. Unless otherwise specified, the processor (110) in the present disclosure may mean a set of one or more processors (110).

[0042] The processor (110) may be implemented as a computer or a similar device according to hardware, software, or a combination thereof. In terms of hardware, the processor (110) may be implemented in the form of an electronic circuit that processes electrical signals to perform a control function, and in terms of software, the processor (110) may be implemented in the form of a program that drives the hardware processor (110). Meanwhile, unless otherwise specified in the description to be described below, the operation of the electronic device (100) may be interpreted as being performed under the control of the processor (110). That is, when modules implemented in the electronic device (100) are executed to perform the image data processing method to be described below, the modules may be interpreted as controlling the processor (110) to perform the following operations of the electronic device (100).

[0043] The memory (120) is a hardware that stores various data processed within the electronic device (100), and the memory (120) can temporarily or semi-permanently store data processed and data to be processed in the electronic device (100). For example, data related to an operating system (OS) for driving the electronic device (100) may be stored in the memory (120) of the electronic device (100). The memory (120) may include a random access memory (RAM) such as a dynamic random access memory (DRAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a CD-ROM, a Blu-ray or other optical disk storage, a hard disk drive (HDD), a solid state drive (SSD), or a flash memory. This memory (120) may be provided as a built-in type or a detachable type. Additionally, the memory (120) can store instructions for the operation of the processor (110). Unless otherwise specified, the memory (120) in the present disclosure may mean a set of one or more memories (120).

[0044] The artificial intelligence accuracy improvement method of the present disclosure performed by the electronic device (100) of FIG. 1 may also be implemented by a non-transitory computer-readable storage medium (or non-transitory recording medium) that is readable by a computer for its operation. The artificial intelligence accuracy improvement method may be implemented by a software module or an algorithm, and may be stored on a computer-readable recording medium as computer-readable codes or program instructions that are executable on a processor (110). Here, the computer-readable recording medium includes a magnetic storage medium (e.g., a read-only memory (ROM), a random-access memory (RAM), a floppy disk, a hard disk, etc.) and an optical reading medium (e.g., a CD-ROM, a Digital Versatile Disc (DVD)). The computer-readable recording medium may be distributed to computer systems connected to a network, so that the computer-readable codes may be stored and executed in a distributed manner. The medium is computer-readable, can be stored in memory (120), and can be executed by a processor (110).

[0045] An electronic device (100) according to one embodiment may further include a display (not shown). The display may visually provide various information to the outside under the control of the processor (110). The display may include a touch circuit configured to detect a user's touch or a sensor circuit configured to measure the intensity of a force generated by a touch.

[0046] The electronic device (100) of FIG. 1 can operate to design an AI based on a Monte Carlo simulation algorithm based on the hardware configuration as described above, set a region of interest for executing a local Monte Carlo simulation within a vulnerable region where the AI ​​is determined to be vulnerable among the application regions of the AI, execute a local Monte Carlo simulation for the region of interest based on the AI, and check information on the improved accuracy of the AI ​​based on the execution of the local Monte Carlo simulation for the region of interest. Here, the AI ​​may be an AI linked to the electronic device (100) so as to provide an operation result through the execution of the Monte Carlo simulation, and the improved accuracy of the AI ​​may be determined based on the fact that the deep learning data acquired through the execution of the local Monte Carlo simulation for the region of interest is additionally reflected in the deep learning data already acquired through the execution of the Monte Carlo simulation for the application region. In other words, the improved accuracy of the AI ​​in Fig. 1 can be judged based on the point set to compare cases where deep learning data acquired through execution of a local Monte Carlo simulation for the area of ​​interest is additionally reflected in the deep learning data acquired through execution of a Monte Carlo simulation for the application area, and cases where it is not reflected.

[0047] AI based on Monte Carlo simulation algorithms suffers from performance limitations when the AI ​​is applied to a large area, such as a "very large number of cases." To address this, the present disclosure identifies vulnerable areas within the AI ​​application area based on Monte Carlo simulation algorithms and sets a subset of the vulnerable areas as a region of interest (ROI). This allows for local Monte Carlo simulations to be performed within the region of interest, thereby improving the accuracy of AI based on Monte Carlo simulation algorithms.

[0048] Regarding the execution of local Monte Carlo simulations, it is necessary to view the overall AI design from the perspective of focusing on and solving a specific area of ​​interest in a particular problem (e.g., a local life-or-death situation in Go, or an area with frequent medical error judgments in medical diagnosis) in a way that is similar to the way existing human experts think.

[0049] In one embodiment, since the Go AI is vulnerable to a water battle in the early central area, rather than setting it to simulate all cases in the entire 19*19 line area, which is the standard Go size, from the initial AI design stage, the vulnerable area is determined, a subset of the area of ​​interest among the determined vulnerable areas is set, the simulation area is limited to the area of ​​interest, and the simulation using the local MCTS algorithm is intensively executed within this limited range, thereby increasing the accuracy of the AI. For example, the local MCTS simulation can be intensively executed only within the central 10*10 line area among the entire 19*19 line area.

[0050] Compared to N*N lines, (N+2)*(N+2) lines of Go, which add one line each to the top, bottom, left, and right, exponentially increase the total number of arrangement cases compared to the existing N*N lines of Go, as 4N+4 stones are added. The 19*19 line of Go, which is currently the most commonly used, is very large for human professional players considering the overall number of cases. The search space size based on 19*19 lines is calculated to be a whopping 10 to the 360th power, so even existing high-performance supercomputers, which calculate all cases, are still woefully inadequate for calculating all of them. The fact that Go, a representative two-player perfect information zero-sum game, is the only board game that a computer could not beat a human world champion until the 20th century proves this.

[0051] The probability that AI has deep learning data on a concentrated battle of the Great Dane, limited to a portion of M*M lines (N > M) rather than the entire N*N line area, is a very rare probability in terms of information entropy according to the MCTS algorithm principle. Therefore, deep learning data resulting from the execution of MCTS simulations on a portion of M*M lines may become a weakness that AI could not learn in advance. Conversely, if deep learning data on a portion of M*M lines (N > M), which is relatively easy to obtain due to the small amount of computation, is utilized, and if the deep learning data on a portion of M*M lines is reflected and substituted into the deep learning data on the entire N*N line area, just like the partial life-or-death thinking of human experts, and managed so that AI solves a given problem, the accuracy and efficiency of AI can be increased.

[0052] In one embodiment, when the AI ​​that is linked to the electronic device (100) and executes a simulation is a Go AI, an M*M line region of interest may be set within a central portion of the entire N*N line Go board area where the AI ​​is determined to be vulnerable (N > M). In this case, deep learning data acquired through execution of a local Monte Carlo simulation for the M*M line region of interest may be additionally reflected in the deep learning data already acquired through execution of a Monte Carlo simulation for the entire N*N line Go board area, thereby improving the accuracy of the AI. According to an embodiment, instead of utilizing only the deep learning data acquired through execution of a Monte Carlo simulation for the entire N*N line Go board area as in the past, deep learning data acquired through execution of a local Monte Carlo simulation for the central M*M line area determined to be a weakness may also be additionally or separately utilized for the operation of the AI. The electronic device (100) can determine and confirm the improved accuracy of the Baduk AI by comparing cases in which deep learning data obtained through execution of a local Monte Carlo simulation as described above is reflected and cases in which it is not.

[0053] In one embodiment, when the AI ​​executing a simulation in connection with the electronic device (100) is a medical diagnosis AI, a medical diagnosis area of ​​interest determined to have a high medical error judgment rate by the AI ​​among the medical diagnosis areas may be set. In this case, deep learning data acquired through execution of a local Monte Carlo simulation for the medical diagnosis area of ​​interest may be additionally reflected in the deep learning data already acquired through execution of a Monte Carlo simulation for the medical diagnosis area, thereby improving the accuracy of the AI. According to an embodiment, rather than utilizing only the deep learning data acquired through execution of a Monte Carlo simulation for the entire medical diagnosis area as in the past, deep learning data acquired through execution of a local Monte Carlo simulation for the medical diagnosis area of ​​interest determined to be a weakness due to a high medical error judgment rate may also be additionally or separately utilized for the operation of the AI. The electronic device (100) can determine and confirm the improved accuracy of the medical diagnosis AI by comparing cases in which the deep learning data acquired through execution of a local Monte Carlo simulation as described above is reflected with cases in which it is not.

[0054] Meanwhile, AI can be designed to be applicable to other algorithms, not just the Monte Carlo simulation algorithm. The Monte Carlo algorithm, a sampling method used for existing machine learning, theoretically possesses randomness, which can lead to homogeneity as the number of samples increases. However, the classical Monte Carlo simulation calculates the proportion of points occupying the target area among the entire area by randomly selecting points and then randomly plotting them. This makes it somewhat incompatible with AI systems that prioritize efficiency. The Monte Carlo method was originally adopted by the mathematician Ulam during the development of the hydrogen bomb, and while it can reflect randomness due to the characteristics of physical phenomena, such randomness can lead to inefficiency in machine learning. The ultimate goal of randomization in Monte Carlo algorithms is homogenization. Strictly speaking, true randomness in Monte Carlo algorithms does not exist; instead, it is a weaker form of pseudo-randomness. Furthermore, due to its inaccuracy, it suffers from inefficiency, requiring numerous trials and resources based on the law of large numbers. While pure randomness might be meaningful as a sample, accurately estimating proportions that conform to the law of large numbers requires numerous trials. Furthermore, in some cases, it requires not only plotting points but also implementing detailed rules that follow from those points, consuming significant time and resources. Therefore, an improved algorithm, rather than the classical Monte Carlo algorithm, could be considered. Instead, it acknowledges the limitations of randomness and focuses on the ultimate goal of randomization: homogenization. Similarly, a Latin Hypercube sampling method, a proportion estimation technique based on appropriate lattice points and grid cells that is easy to implement and guarantees uniformity, could be applied to this disclosure.

[0055] Lattice points can be a new approach different from the Monte Carlo algorithm because they are simpler rules and ensure uniformity rather than randomness, and the estimation of associated values ​​can also be more accurate and efficient. The application of the Latin hypercube sampling method can be understood as having been inspired by the fact that it is most efficient to ensure uniformity by finding the minimum number of component members on a coordinate plane such as a checkerboard when calculating the area ratio. Compared to general Monte Carlo simulations, the Latin hypercube sampling method is known to have the advantage of having a homogeneous sample distribution and being able to pursue the same level of statistical accuracy with a small number of samples, and is therefore menu-ized in existing statistical programs. However, since there has been no case of applying it to increase the accuracy of AI, the embodiment according to the present disclosure can be interpreted as a new attempt.

[0056] In one embodiment, when the AI ​​applies the Monte Carlo simulation algorithm, the Latin hypercube algorithm may be set to be applicable instead of the Monte Carlo simulation algorithm depending on conditions related to uniformity and / or efficiency. Specifically, when the AI ​​applies the Monte Carlo simulation algorithm, if the uniformity confirmed for each detailed area with respect to the AI's operations and results targeting the application area, vulnerable area, and / or area of ​​interest is below a certain level, or if the computational efficiency of the AI ​​targeting the application area, vulnerable area, and / or area of ​​interest is determined to be below a certain level, the AI ​​may apply the Latin hypercube algorithm instead of the Monte Carlo simulation algorithm. In this case, the AI ​​may perform operations according to various embodiments of the present disclosure described above based on the Latin hypercube algorithm, and the electronic device (100) may also check information about the improved accuracy of the AI ​​based on the Latin hypercube algorithm.

[0057] In one embodiment, in order to fundamentally prevent frequent errors of AI that appear in a specific area among the application areas of AI, if the inference value of AI is determined to correspond to a state such as a clearly known bug, an incorrect inference result that violates the regulations, etc., the AI ​​may be forced to ignore the inference process and result according to the specific area and determine the inference process and result through its own algorithm set in advance in the AI ​​to prevent mistakes, thereby further improving the accuracy of the AI ​​based on the Monte Carlo simulation algorithm. In other words, if the AI ​​is determined to correspond to an application error state or an inference error state in a specific area among the application areas of AI, the application result of the AI ​​for the specific area may be excluded based on the processing algorithm set in the AI ​​and only the application result of the AI ​​for the remaining areas excluding the specific area among the application areas may be obtained.

[0058] In one embodiment, when the electronic device (100) determines and confirms the improved accuracy based on the local Monte Carlo simulation of the AI ​​for the region of interest, the electronic device (100) may store the size of the region of interest among the vulnerable regions and the improved accuracy of the AI ​​in the case where the local Monte Carlo simulation is performed for the region of interest of the corresponding size as information. In addition, the electronic device (100) may change the size of the region of interest step by step according to a certain degree within the vulnerable region, and store the information as described above for each size of the region of interest and the improved accuracy of the AI ​​corresponding to the size of each region of interest, thereby determining the size of the region of interest within the vulnerable region with the most improved accuracy based on this.

[0059] Taking Go AI as an example, when the area where AI is vulnerable among the entire Go board area of ​​19*19 lines, which is the standard specification, is the central area of ​​10*10 lines, the electronic device (100) stores 1) information including the improved accuracy of AI that is confirmed when deep learning data acquired by executing a local Monte Carlo simulation for a 5*5 line area of ​​interest and the 5*5 line area of ​​interest are reflected, 2) information including the improved accuracy of AI that is confirmed when deep learning data acquired by executing a local Monte Carlo simulation for a 6*6 line area of ​​interest and the 6*6 line area of ​​interest are reflected, ..., 3) information including the improved accuracy of AI that is confirmed when deep learning data acquired by executing a local Monte Carlo simulation for a 10*10 line area of ​​interest and the 10*10 line area of ​​interest are reflected, and based on the stored information, stores information including the improved accuracy of AI that is confirmed when deep learning data acquired by executing a local Monte Carlo simulation for a 5*5 line area of ​​interest and the 6*5 line area of ​​interest. Among the lines, ..., 10*10 lines, the size of the optimal region of interest with the highest improved accuracy of the AI ​​can be determined. Afterwards, when the electronic device (100) needs to run a Monte Carlo simulation targeting the entire 19*19 line checkerboard area in other similar situations, the performance of the AI ​​can be improved by running a local Monte Carlo simulation based on the previously determined optimal size region of interest and managing the deep learning data acquired through this to be reflected in the operation of the AI.

[0060] In summary, the electronic device (100) can set multiple regions of interest with different sizes within a vulnerable region, and when deep learning data obtained by executing a local Monte Carlo simulation for each region of interest included in the multiple regions of interest is reflected, the improved accuracy of the AI ​​determined for each can be confirmed, and the region of interest of the optimal size with the highest accuracy can be confirmed. The electronic device (100) can confirm the region of interest of the optimal size within the vulnerable region in terms of accuracy, and can utilize the related information for AI operation in similar situations in the future.

[0061] Meanwhile, depending on the size of the vulnerable area, there may be a large difference in the size to which the region of interest can be set. Taking the aforementioned Baduk AI as an example, from the perspective of AI simulation, it can be understood that there is a significant difference in the size of the region of interest when the region of interest is set to 5*5 lines and when it is set to 10*10 lines. Even if the AI ​​may have improved accuracy when the region of interest is set to 10*10 lines when executing a local Monte Carlo simulation than when the region of interest is set to 5*5 lines, the 10*10 line region of interest generates a much larger computational load than the 5*5 line region of interest. Therefore, it may be difficult to conclude that the 10*10 line region of interest is an optimally sized region of interest compared to the 5*5 line region of interest in terms of AI performance that considers computational load in addition to accuracy.

[0062] Therefore, when determining the optimal size of the region of interest for AI to perform a local Monte Carlo simulation, not only the improved accuracy confirmed corresponding to the size of each region of interest as described above, but also the computational load confirmed corresponding to the size of each region of interest can be reflected. That is, when determining the optimal size of the region of interest for AI performance among regions of interest that can be set to various sizes within the vulnerable region, the electronic device (100) can determine the optimal size of the region of interest by considering the improved accuracy confirmed for each region of interest and the degree of computational load for each region of interest.

[0063] First, the electronic device (100) can express each improved accuracy confirmed in response to the size of each region of interest as a percentage. In addition, the electronic device (100) can determine each degree of computational load confirmed for each region of interest of each size as a percentage between 0% and 100%, assuming that there is no computational load as 0% and that the AI ​​cannot operate due to the computational load as 100%. Based on this, the electronic device (100) can determine an optimally sized region of interest by synthesizing the percentage of each improved accuracy confirmed in response to the size of each region of interest and the percentage of each degree of computational load confirmed in response to the size of each region of interest. The degree of computational load can be set to be reflected as a negative (-) value when determining the optimal size of the region of interest, since the higher the percentage, the greater the computational load, which hinders the performance of the AI. Furthermore, the electronic device (100) can flexibly set the accuracy weight for improved accuracy and the computational load weight for the degree of computational load, taking into account whether to focus on accuracy in the performance of the AI ​​or to focus on preventing computational load, so that the optimal size of the region of interest can be determined differently according to the weights.

[0064] The above example for determining the region of interest of the optimal size can be expressed mathematically as [Mathematical Formula 1] below.

[0065] [Mathematical Formula 1]

[0066] argmax((f(x))

[0067] f(x) = *Improved Accuracy (%) - *Computational load level (%)

[0068] (x = size of each region of interest, = accuracy weight, = Operational load weight)

[0069] According to [Mathematical Formula 1], the improved accuracy and computational load can be checked for each size of the region of interest, and the f(x) value can be checked by multiplying each weight, and the size of the region of interest with the largest f(x) value can be determined. In terms of accuracy and computational load, the electronic device (100) checks the region of interest with the size determined through [Mathematical Formula 1] as the region of interest with the optimal size within the vulnerable region, and the related information can be utilized for AI operation in similar situations in the future.

[0070] It is obvious that each piece of information or data generated in the process of an electronic device (100) performing a method to improve artificial intelligence accuracy can be combined in various forms.

[0071] As examined above, the existing Monte Carlo simulation algorithm-based AI had a vulnerable part due to the problem that the number of cases for the target service was very large, but according to the present disclosure, rather than simulating all cases for the entirety, the simulation area is reduced to a vulnerable area, and in particular, by setting a subset range of the area of ​​interest and then executing a local Monte Carlo simulation algorithm within this limited range, the accuracy of the AI ​​can be increased.

[0072] Although all components constituting the embodiments disclosed in this document have been described as being combined or operating in combination as one, the embodiments disclosed in this document are not necessarily limited to such embodiments. That is, within the scope of the purpose of the embodiments disclosed in this document, all of the components may be selectively combined and operated one or more times.

[0073] In addition, terms such as "include," "comprise," or "have" described above, unless specifically stated otherwise, mean that the corresponding component can be included, and therefore should be interpreted to include other components rather than excluding other components. All terms, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art to which the embodiments disclosed in this document belong, unless otherwise defined. Commonly used terms, such as terms defined in a dictionary, should be interpreted to be consistent with the contextual meaning of the relevant technology, and shall not be interpreted in an idealized or overly formal sense, unless explicitly defined in this document.

[0074] The above description is merely an example of the technical idea disclosed in this document, and those skilled in the art to which the embodiments disclosed in this document pertain may make various modifications and variations without departing from the essential characteristics of the embodiments disclosed in this document. Therefore, the embodiments disclosed in this document are not intended to limit the technical idea of ​​the embodiments disclosed in this document, but to explain it, and the scope of the technical idea disclosed in this document is not limited by these embodiments. The scope of protection of the technical idea disclosed in this document should be interpreted by the claims below, and all technical ideas within a scope equivalent thereto should be interpreted as being included in the scope of rights of this document.

Claims

1. In a method for improving the accuracy of artificial intelligence (AI) of electronic devices, Steps to design AI based on Monte-Carlo simulation algorithm; A step of setting a region of interest for executing a local Monte Carlo simulation within a vulnerable region where the AI ​​is determined to be vulnerable among the application regions of the AI; Based on the above AI, a step of executing a local Monte Carlo simulation for the region of interest; and A step of identifying information about the improved accuracy of the AI ​​based on the execution of a local Monte Carlo simulation for the region of interest, The above improved accuracy is judged based on the fact that deep learning data obtained through execution of a local Monte Carlo simulation for the area of ​​interest is additionally reflected in the deep learning data obtained through execution of a Monte Carlo simulation for the application area. How to improve AI accuracy.

2. In paragraph 1, The above AI is set to be able to apply the Latin Hypercube algorithm instead of the Monte Carlo simulation algorithm based on conditions related to uniformity and efficiency. How to improve AI accuracy.

3. In paragraph 1, When the above AI responds to the Baduk AI, an M*M line area of ​​interest is set within the central part area where the AI ​​is judged to be vulnerable among the entire N*N line area of ​​the Go board. The above improved accuracy is judged based on the fact that deep learning data acquired through execution of a local Monte Carlo simulation for the M*M line area of ​​interest is additionally reflected in the deep learning data acquired through execution of a Monte Carlo simulation for the entire N*N line checkerboard area, The above M is smaller than the above N, How to improve AI accuracy.

4. In paragraph 1, When the above AI corresponds to a medical diagnosis AI, a medical diagnosis area of ​​interest is set in which the medical error judgment rate by the AI ​​is judged to be high among the medical diagnosis areas. The above improved accuracy is determined based on the fact that deep learning data obtained through execution of a local Monte Carlo simulation for the medical diagnosis area of ​​interest is additionally reflected in the deep learning data obtained through execution of a Monte Carlo simulation for the medical diagnosis area of ​​interest. How to improve AI accuracy.

5. In paragraph 1, If the AI ​​is determined to be in an application error state for a specific area among the above application areas, the application result of the AI ​​for the specific area is excluded based on the processing algorithm preset in the AI, and only the application result of the AI ​​for the remaining areas excluding the specific area among the application areas is set to be obtained. How to improve AI accuracy.

6. In paragraph 1, The above AI accuracy improvement method is, A step of setting multiple areas of interest having different sizes within the above vulnerable area; For each region of interest included in the plurality of regions of interest, a step of confirming each improved accuracy judged for the AI ​​based on the execution of a local Monte Carlo simulation for each region of interest; and Further comprising a step of setting information including the size of each of the above areas of interest and the improved accuracy of each of the above areas of interest. How to improve AI accuracy.

7. In paragraph 6, The above AI accuracy improvement method is, A step of checking the degree of each computational load determined when executing a local Monte Carlo simulation for each region of interest included in the above multiple regions of interest; and Further comprising a step of determining a region of interest of an optimal size for the performance of the AI ​​among the plurality of regions of interest based on each of the improved accuracy and each of the computational load levels. How to improve AI accuracy.

8. In an electronic device that performs a method for improving the accuracy of artificial intelligence (AI), processor; and Contains one or more memories that store one or more instructions, The one or more instructions, when executed, cause the processor to: Steps to design AI based on Monte-Carlo simulation algorithm; A step of setting a region of interest for executing a local Monte Carlo simulation within a vulnerable region where the AI ​​is determined to be vulnerable among the application regions of the AI; Based on the above AI, a step of executing a local Monte Carlo simulation for the region of interest; and Controlling the processor to perform a step of verifying information about the improved accuracy of the AI ​​based on the execution of a local Monte Carlo simulation for the region of interest; The above improved accuracy is judged based on the fact that deep learning data obtained through execution of a local Monte Carlo simulation for the area of ​​interest is additionally reflected in the deep learning data obtained through execution of a Monte Carlo simulation for the application area. Electronic devices.

9. In paragraph 8, The above AI is set to be able to apply the Latin Hypercube algorithm instead of the Monte Carlo simulation algorithm based on conditions related to uniformity and efficiency. Electronic devices.

10. In paragraph 8, When the above AI responds to the Baduk AI, an M*M line area of ​​interest is set within the central part area where the AI ​​is judged to be vulnerable among the entire N*N line area of ​​the Go board. The above improved accuracy is judged based on the fact that deep learning data acquired through execution of a local Monte Carlo simulation for the M*M line area of ​​interest is additionally reflected in the deep learning data acquired through execution of a Monte Carlo simulation for the entire N*N line checkerboard area, The above M is smaller than the above N, Electronic devices.

11. In paragraph 8, When the above AI corresponds to a medical diagnosis AI, a medical diagnosis area of ​​interest is set in which the medical error judgment rate by the AI ​​is judged to be high among the medical diagnosis areas. The above improved accuracy is determined based on the fact that deep learning data obtained through execution of a local Monte Carlo simulation for the medical diagnosis area of ​​interest is additionally reflected in the deep learning data obtained through execution of a Monte Carlo simulation for the medical diagnosis area of ​​interest. Electronic devices.

12. In paragraph 8, If the AI ​​is determined to be in an application error state for a specific area among the above application areas, the application result of the AI ​​for the specific area is excluded based on the processing algorithm preset in the AI, and only the application result of the AI ​​for the remaining areas excluding the specific area among the application areas is set to be obtained. Electronic devices.

13. In paragraph 8, The above processor, A step of setting multiple areas of interest having different sizes within the above vulnerable area; For each region of interest included in the plurality of regions of interest, a step of confirming each improved accuracy judged for the AI ​​based on the execution of a local Monte Carlo simulation for each region of interest; and Further performing the step of setting information including the size of each of the above areas of interest and the above improved accuracy, Electronic devices.

14. In paragraph 13, The above processor, A step of checking the degree of each computational load determined when executing a local Monte Carlo simulation for each region of interest included in the above multiple regions of interest; and Based on each of the above improved accuracy and each of the above computational load levels, a step of determining an area of ​​interest of an optimal size for the performance of the AI ​​among the plurality of areas of interest is further performed. Electronic devices.

Images