Method for generating noise for neural network model by using scoring-based filtering and noise distribution model

Abstract

Disclosed is a method for generating noise for a neural network model by using scoring-based filtering and a noise distribution model. The method according to an embodiment of the present disclosure may comprise the steps of: acquiring a condition vector for a neural network model; generating noise to be input to the neural network model; inputting the noise and the condition vector to a noise score inference model; identifying whether a score output from the noise score inference model is greater than a threshold; inputting the noise and the condition vector to the neural network model on the basis of identifying that the score is greater than the threshold; and regenerating noise on the basis of identifying that the score is not greater than the threshold.

Description

Method for generating noise for neural network models using score-based filtering and noise distribution models

[0001] The present disclosure relates to a neural network model using noise, and more specifically, to a method for generating noise for a neural network model using a score-based filtering and a noise distribution model.

[0002] Neural network models, or Artificial Intelligence (AI) models, are mathematical models modeled after the structure of neurons in the human brain, capable of receiving data input and performing prediction or classification tasks. Neural network models can learn data and perform predictions through multiple layers, such as input layers, hidden layers, and output layers, which contain multiple nodes or neurons.

[0003] Generative AI models can be neural network models or AI models specialized in generating new data from given data. Generative AI models can generate creative and novel samples based on learned distributions. By learning from existing data and based on the learned distributions, Generative AI models can generate various forms of output, such as text, images, or music.

[0004] A method according to one embodiment of the present disclosure may include: obtaining a condition vector for a neural network model; generating noise to be input to the neural network model; inputting the noise and the condition vector to a noise score inference model; identifying whether a score output from the noise score inference model is greater than a threshold; inputting the noise and the condition vector to the neural network model based on identifying that the score is greater than the threshold; and regenerating noise based on identifying that the score is not greater than the threshold.

[0005] A computer-readable recording medium according to one embodiment of the present disclosure may record a program for performing at least one of the methods according to one embodiment of the present disclosure at least partially on a computer.

[0006] An electronic device according to one embodiment of the present disclosure may include at least one processor comprising a processing circuit, and one or more storage media and a memory for storing one or more instructions. When the one or more instructions are executed individually by the at least one processor, the electronic device may: acquire a condition vector for a neural network model; generate noise to be input to the neural network model; input the noise and the condition vector to a noise score inference model; identify whether a score output from the noise score inference model is greater than a threshold; input the noise and the condition vector to the neural network model based on identifying that the score is greater than the threshold; and regenerate the noise based on identifying that the score is not greater than the threshold.

[0007] FIG. 1 illustrates an exemplary method for generating an output from a neural network model using score-based filtered noise according to one embodiment of the present disclosure.

[0008] FIG. 2 illustrates an exemplary block diagram of an electronic device including a noise score inference model and a neural network model according to one embodiment of the present disclosure.

[0009] FIG. 3 exemplarily illustrates a method for generating an output from a neural network model using score-based filtered noise according to one embodiment of the present disclosure.

[0010] FIG. 4 illustrates an exemplary method for updating a noise score inference model according to one embodiment of the present disclosure.

[0011] FIG. 5 exemplarily illustrates a method for generating an output from a neural network model based on noise generated using a noise distribution model according to one embodiment of the present disclosure.

[0012] FIG. 6 exemplarily illustrates a flowchart of a method for acquiring noise to be input to a neural network model based on a condition vector according to one embodiment of the present disclosure.

[0013] FIG. 7 illustrates an exemplary block diagram of an electronic device according to one embodiment of the present disclosure.

[0014] FIG. 8 illustrates an exemplary flowchart of a method for providing an output from a neural network model to a user in response to a condition vector input by a user, according to one embodiment of the present disclosure.

[0015] FIG. 9 illustrates, in accordance with one embodiment of the present disclosure, a user interface displayed on the display device of FIG. 7.

[0016] FIG. 10 illustrates an exemplary block diagram of an electronic device according to one embodiment of the present disclosure.

[0017] FIG. 11 illustrates an exemplary flowchart of a method for providing an output from a neural network model to a user in response to a condition vector input by a user, according to one embodiment of the present disclosure.

[0018] The terms used in the embodiments of this specification have been selected to be as widely used as possible, taking into account the functions of the present disclosure; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the description section of the relevant embodiments. Therefore, terms used in this specification should be defined not merely by their names, but based on their meanings and the overall content of the present disclosure.

[0019] The terms used in this disclosure are used merely to describe specific embodiments and are not intended to limit the scope of other embodiments. A singular expression may include a plural expression unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, may have the same meaning as generally understood by those skilled in the art described in this disclosure. Terms used in this disclosure that are defined in a general dictionary may be interpreted as having the same or similar meaning as they have in the context of the relevant technology, and are not to be interpreted in an ideal or overly formal sense unless explicitly defined in this disclosure. In some cases, even terms defined in this disclosure are not to be interpreted to exclude the embodiments of this disclosure.

[0020] In the various embodiments of the present disclosure described below, a hardware-based approach is described as an example. However, since the various embodiments of the present disclosure include techniques using both hardware and software, the various embodiments of the present disclosure do not exclude a software-based approach.

[0021] Singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, may have the same meaning as generally understood by those skilled in the art as described in this specification.

[0022] Throughout this disclosure, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "…part," "…module," etc., as used in this specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0023] 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. Instead, in some contexts, the expression “system configured to” may mean that the system is “capable of” in conjunction with other devices or components. For example, the phrase “processor configured to perform A, B, and C” may mean a dedicated processor for performing the said operations (e.g., an embedded processor), or a generic-purpose processor (e.g., a CPU or an application processor) capable of performing said operations by executing one or more software programs stored in memory.

[0024] In addition, when a component is described in the present disclosure as being "connected" or "connected" to another component, it should be understood that the component may be directly connected to or directly connected to the other component, but unless otherwise specifically stated, it may also be connected or connected through another component in between.

[0025] Additionally, in this disclosure, expressions such as "greater than" or "less than" may be used to determine whether a specific condition is satisfied or fulfilled; however, this is merely for the purpose of expressing an example and does not exclude descriptions such as "greater than" or "less than." Conditions described as "greater than" may be replaced with "greater than," conditions described as "less than" may be replaced with "less than," and conditions described as "greater than and less than" may be replaced with "greater than and less than."

[0026] FIG. 1 illustrates an exemplary method for generating an output from a neural network model (120) using score-based filtered noise according to one embodiment of the present disclosure.

[0027] Referring to FIG. 1, a noise score inference model (110) may receive one or more noises and a condition vector. The noise score inference model (110) may infer (or estimate) the score of each noise based on the condition vector. Based on the scores inferred by the noise score inference model (110), one or more noises may be filtered. For example, among one or more noises, only the noise(s) whose inferred score exceeds a threshold may be input to the neural network model (120). The noise(s) whose inferred score exceeds the threshold may be understood as score-based filtered noise(s). The neural network model (120) may receive score-based filtered noise(s) and a condition vector. Based on the score-based filtered noise(s), the neural network model (120) may infer an output that satisfies the requirements, restrictions, or contexts indicated by the condition vector.

[0028] In some embodiments of the present disclosure, the score inferred by the noise score inference model (110) may correspond to the inferred (or estimated) quality of the output generated by the neural network model (120) when any noise is input to the neural network model (120). For example, when the neural network model (120) outputs an image, the score inferred by the noise score inference model (110) may indicate a score for indicators for evaluating the image, such as whether the image generated by the neural network model (120) satisfies the requirements, and whether the composition and content of the image are appropriate. When the neural network model (120) outputs text, the score inferred by the noise score inference model (110) may indicate a score for indicators for evaluating the text, such as whether the text generated by the neural network model (120) satisfies the requirements, and whether the structure, grammar, and words of the text are appropriate.

[0029] In some embodiments of the present disclosure, a 'condition vector' may refer to input data or input representations used to control the execution of a neural network model based on specific conditions or information. A condition vector may provide additional degrees or constraints to guide the output of the neural network model. For example, a condition vector may be input to a neural network model to use the neural network model to generate content that meets specific requirements or contexts. A condition vector may be a structured representation that encodes conditions or contextual information provided by a user. In one embodiment, a condition vector may be generated from user input, such as natural language prompts or category labels. In one embodiment, a condition vector may be generated by converting text data into an embedding vector or by separate preprocessing.

[0030] In some embodiments of the present disclosure, ‘noise’ may refer to random data or a random vector input to the neural network model (120) for use in inference of the neural network model (120). The neural network model (120) may regard the given noise as a point in a high-dimensional latent space and generate new data samples based on the given noise. The noise may also be referred to as a ‘noise vector’.

[0031] In one embodiment, the neural network model (120) may be a generative model or a generative artificial intelligence (AI) model that produces various forms of data such as text, images, video, or audio. For example, the neural network model (120) may learn basic patterns and structures from training data and generate new data based on inputs such as condition vectors. The neural network model (120) may use score-based filtered noise to generate an output that satisfies the requirements indicated by a given condition vector. The condition vector controls the output direction of the neural network model (120), and the noise can ensure diversity in the output of the neural network model (120). In one embodiment, the neural network model (120) may be implemented based on a Generative Adversarial Network (GAN), a Variational Autoencoder (VAE), a diffusion model, an Autoregressive model, a flow-based model, a transformer-based model, or any combination thereof.

[0032] Even if the same condition vector is used, the quality of the output of the neural network model (120) may change due to noise. Meanwhile, if the output of the neural network model (120) is generated using a predefined noise set to ensure high quality, the diversity of the neural network model (120) may be reduced. In addition, the conditions input by the user may vary, and therefore the noise optimized for each condition may differ.

[0033] In the embodiment illustrated in FIG. 1, noise acquired to be input into the neural network model (120) can be filtered based on a scoring method. The noise score inference model (110) can be pre-trained based on a number of learning conditions and learning noises. For example, the noise score inference model (110) can be trained to infer the score of each of the learning noises for a single learning condition. The noise score inference model (110) can score one or more noises based on a condition vector. Among the one or more noises, only the noise(s) whose inferred score is greater than a threshold can be input into the neural network model (120). Noise(s) whose inferred score is not greater than the threshold can be deleted, eliminated, or discarded. Accordingly, the quality of the output generated by the neural network model (120) can be improved.

[0034] FIG. 2 illustrates an exemplary block diagram of an electronic device (200) including a noise score inference model (110) and a neural network model (120) according to one embodiment of the present disclosure.

[0035] Referring to FIG. 2, the electronic device (200) may include a noise generation module (210), a noise filtering module (220), an inference module (230), and a model update module (240). The configuration of the electronic device (200) shown in FIG. 2 is merely an example, and the examples of electronic devices performing an embodiment of the present disclosure are not limited to the configuration shown in FIG. 2. In one embodiment, one or more of the configurations shown in FIG. 2 may be deleted or changed, or a configuration not shown in FIG. 2 may be added to the electronic device (200).

[0036] The noise generation module (210) can generate one or more noises to be input into the neural network model (120). The noise generation module (210) may include a noise distribution model (212) that outputs one or more noises. The noise distribution model (212) may be a model representing a distribution for generating noise or an algorithm for generating noise.

[0037] In one embodiment, the noise distribution model (212) may be based on a parametric model or a non-parametric model. For example, one or more functions (or kernels or models) representing the distribution of noise and parameters for each function may be determined while defining the noise distribution model (212). The number of parameters for each function model may or may not be limited. The noise distribution model (212) may be used to sample one or more noises to be input to a neural network model (120) based on a given condition vector. In one embodiment, the noise distribution model (212) may be implemented based on a parametric or non-parametric model such as a neural network, a Gaussian process, Kernel Density Estimation (KDE), a Support Vector Machine (SVM), or K-nearest neighbors.

[0038] In one embodiment, the noise distribution model (212) may correspond to a Gaussian distribution. The noise generated by the noise distribution model (212) may correspond to Gaussian noise. For example, based on the noise distribution model (212), random noise based on a Gaussian distribution may be generated.

[0039] The noise filtering module (220) can infer a score of the noise generated by the noise generation module (210) and filter the noise based on the inferred score. The noise filtering module (220) can infer a score of the noise generated by the noise generation module (210) and filter the noise based on the inferred score. The noise filtering module (220) may include a noise score inference model (110) and a score-based filter (222). The noise score inference model (110) can infer a score for the noise generated by the noise generation module (210).

[0040] A scoring-based filter (222) can filter noise generated by a noise generation module (210) based on a score inferred by a noise score inference model (110). For example, the scoring-based filter (222) can identify (or determine) whether the score inferred by the noise score inference model (110) is greater than a predefined threshold. Based on identifying (or determining) that the score inferred by the noise score inference model (110) is greater than a predefined threshold, the scoring-based filter (222) can provide the corresponding noise to the inference module (230). Based on identifying that the score inferred by the noise score inference model (110) is not greater than a predefined threshold, the scoring-based filter (222) can delete, remove, or discard the corresponding noise. Based on identifying that some or all of the inferred scores of the noises generated by the noise generation module (210) are not greater than a predefined threshold, the noise filtering module (220) may request the noise generation module (210) to regenerate the noise.

[0041] The inference module (230) can perform inference based on a given condition vector by executing a neural network model (120). A given condition vector and noise filtered by a noise filtering module (220) can be input into the neural network model (120). The neural network model (120) can generate an output corresponding to the given condition vector using the noise filtered by the noise filtering module (220). The output generated by the neural network model (120) can be provided to the user who input the condition vector.

[0042] The model update module (240) may include an evaluation module (242) and an update module (244). The evaluation module (242) may obtain an evaluation of the output generated by the inference module (230). For example, the evaluation module (242) may request an evaluation of the generated output from a user who has provided a given condition vector. The evaluation module (242) may perform an evaluation of the generated output based on the given condition vector. Based on the evaluation obtained by the evaluation module (242), the update module (244) may update the noise distribution model (212) and / or the noise score inference model (110). For example, based on the evaluation obtained by the evaluation module (242), the update module (244) may modify one or more parameters of the noise distribution model (212) and / or the noise score inference model (110).

[0043] FIG. 3 illustrates an exemplary method for generating an output from a neural network model (120) using score-based filtered noise according to one embodiment of the present disclosure.

[0044] Referring to FIG. 3, the noise generation module (210) of FIG. 2 can sample noise from a noise distribution model (212) to be input to a neural network model (120). For example, the noise generation module (210) can receive a condition vector and generate noise based on the condition vector and the noise distribution model (212). For example, the noise distribution model (212) can correspond to a predefined noise distribution (e.g., a Gaussian distribution), and the noise generation module (210) can extract random noise from this noise distribution model (212). The noise generated by the noise generation module (210) can be transmitted to a noise filtering module (220).

[0045] The noise score inference model (110) of the noise filtering module (220) can infer a score of the noise generated by the noise generation module (210) based on a condition vector. For example, the noise score inference model (110) can estimate a score indicating the quality of the output to be generated by the neural network model (120) based on the noise generated by the noise generation module (210) and the condition vector. The noise score inference model (110) can provide the inferred score of the noise to a scoring-based filter (222). The scoring-based filter (222) can identify whether the score of the noise inferred by the noise score inference model (110) exceeds a predefined threshold. If the score of the inferred noise exceeds the predefined threshold, the scoring-based filter (222) can transmit the noise to the inference module (230). Otherwise, the scoring-based filter (222) can discard the noise.

[0046] In one embodiment, based on identifying that the score of the inferred noise does not exceed a predefined threshold, the score-based filter (222) may request the noise generation module (210) to regenerate the noise. In response to the request to regenerate the noise, the noise generation module (210) may regenerate new noise using the noise distribution model (212) and transmit the regenerated noise to the noise filtering module (220).

[0047] In one embodiment, the noise score inference model (110) may be implemented based on a neural network model. For example, the noise score inference model (110) may be a neural network model having one or more weights that are pre-trained to score how well the output generated by the neural network model (120) based on the given noise matches a given condition vector. The noise score inference model (110) may receive a condition vector and noise, and output a score for the noise.

[0048] In one embodiment, the noise score inference model (110) may be implemented based on a Gaussian process. For example, the noise score inference model (110) may include an average function and one or more kernel functions. The noise score inference model (110) may use the average function and one or more kernel functions to predict the probability of whether the output to be generated by the neural network model (120) based on the given noise will match a given condition vector. The noise score inference model (110) may convert the predicted probability into a score of the noise.

[0049] In one embodiment, the noise score inference model (110) may be implemented based on a discriminative model such as SVM or KNN. The noise score inference model (110) may be trained to classify the output to be generated by the neural network model (120) based on the noise into 'good noise' that matches a given condition vector and 'bad noise' that does not match a given condition vector, and to predict which of the two classes the input noise will be labeled as. The noise score inference model (110) may predict the probability that the output to be generated by the neural network model (120) based on the given noise will match a given condition vector (e.g., the probability that the noise will be labeled as 'good noise') or not match (e.g., the probability that the noise will be labeled as 'bad noise'). The noise score inference model (110) may convert the predicted probability into a score of the noise.

[0050] The inference module (230) can execute the neural network model (120). The inference module (230) can input the condition vector and noise, provided by the noise filtering module (220), into the neural network model (120), where the inference score is greater than a threshold. Based on the condition vector and the noise, the neural network model (120) can output a result that satisfies the requirements indicated by the condition vector.

[0051] In one embodiment, the neural network model (120) may be implemented as a generative AI model. Noise input to the generative AI model can ensure diversity in the output of the generative AI model. By using different noises, the generative AI model can generate different outputs for a given condition vector. Each noise input to the generative AI model can be considered to correspond to a phenomenon (or information, data) in the actual reality. For example, the generative AI model may consider the noise as a point in the potential space and generate data by referencing that point. Therefore, if the noise value is different, the generative AI model can explore a different location in the potential space and, as a result, generate a different result for a given condition vector.

[0052] Due to various reasons, such as limitations in the size of the generative AI model, restrictions on the amount of training data, or limitations on the training data, the quality generated by the generative AI model may be degraded depending on the input noise. For example, the characteristics of the data generated by the generative AI model (image composition; text style, grammar, or words; audio noise, etc.) may be inappropriate. The data generated by the generative AI model may be irrelevant to given conditions. The data generated by the generative AI model may satisfy given conditions but be harmful according to social norms. For example, data generated by the generative AI model may match a given condition vector but be perceived by people as violent or sexually explicit.

[0053] The noise filtering module (220) can filter noise generated by the noise generation module (210). The noise filtering module (220) can infer a score indicating the quality of the result to be generated by the neural network model (120) based on each noise, and filter the noise based on the inferred score. Accordingly, noise having a score greater than a threshold can be provided to the neural network model (120). As a result, while the diversity of the noise is ensured, the quality of the output of the neural network model (120) is improved at the same time, and instability of the output (e.g., inappropriate characteristics, output unrelated to given conditions, or harmful output) can be eliminated.

[0054] FIG. 4 illustrates an exemplary method for updating a noise score inference model (110) according to one embodiment of the present disclosure.

[0055] Referring to FIG. 4, the noise filtering module (220) can provide noise to the inference module (230) having a score inferred by the noise score inference model (110), which is generated based on a condition vector and is greater than a predefined threshold. The inference module (230) can input the noise and the condition vector to the neural network model (120). The neural network model (120) can use the noise to generate an output that matches the given condition vector. The output generated by the neural network model (120) can be transmitted to the model update module (240).

[0056] The model update module (240) can receive a condition vector and the output of the neural network model (120). The model update module (240) can update the noise score inference model (110) based on the condition vector and the output of the neural network model (120). The model update module (240) may include an evaluation module (242) and an update module (244). The evaluation module (242) can evaluate the output of the neural network model (120) based on the condition vector. The evaluation module (242) may include a reward model (402) and / or a feedback request module (404).

[0057] The reward model (402) can quantitatively evaluate the output of the neural network model (120) based on various evaluation criteria, such as goals, quality standards, and / or user preferences, indicated by a condition vector. The reward model (402) evaluates whether the output of the neural network model (120) is good or bad in light of a given condition vector and can convert the evaluation into a quantified score.

[0058] In one embodiment, when the neural network model (120) generates an image, the reward model (402) may evaluate the image generated by the neural network model (120) based on indicators such as how well the image generated by the neural network model (120) matches a condition vector, the quality of the image (e.g., resolution, sharpness, color accuracy), the realism of the image, and whether it satisfies content restriction conditions (e.g., violence, obscenity, or sensitivity). In one embodiment, when the neural network model (120) generates text, the reward model (402) may evaluate the text generated by the neural network model (120) based on indicators such as how well the text generated by the neural network model (120) matches a condition vector, the quality of the text (e.g., grammatical and linguistic accuracy, stylistic consistency, or readability), the accuracy of the content, and whether it satisfies content restriction conditions (e.g., violence, obscenity, or sensitivity).

[0059] The reward model (402) can perform an evaluation of the output generated by the neural network model (120) and provide the evaluation result to the update module (244). The update module (244) can update the noise score inference model (110) based on the evaluation of the reward model (402). In one embodiment, based on the evaluation of the reward model (402), the update module (244) can adjust one or more parameters of the noise score inference model (110) and / or one or more policies of the noise score inference model (110).

[0060] The feedback request module (404) may request feedback on the output of the neural network model (120) from an external device or a user of the electronic device (200). For example, the feedback request module (404) may transmit the output of the neural network model (120) along with a request to evaluate the output of the neural network model (120) to an external device. The feedback request module (404) may provide the output of the neural network model (120) along with a request to evaluate the output of the neural network model (120) to a user of the electronic device (200). In response to a request to evaluate the output of the neural network model (120), the feedback request module (404) may obtain an evaluation of the output of the neural network model (102) from the external device or user. The evaluation of the output of the neural network model (102) may be text including a quantified score, a label for the output of the neural network model (102), and / or a comment for the output of the neural network model (102).

[0061] The update module (244) can update the noise score inference model (110) based on an evaluation of the output obtained by the evaluation module (242). In one embodiment, if the noise score inference model (110) is implemented based on a parametric model such as a neural network model or a regression model, the update module (244) can adjust one or more parameters (e.g., weights) of the noise score inference model (110) based on the evaluation of the evaluation module (242). For example, if the evaluation module (242) assigns a score greater than a predefined threshold to the output of the neural network model (120) or assigns a positive label (e.g., a label corresponding to 'good output') to the output of the neural network model (102), the update module (244) can adjust one or more parameters of the noise score inference model (110) so that the noise score inference model (110) assigns a higher score to the noise. In one embodiment, the update module (244) can update one or more parameters of the noise score inference model (110) using reinforcement learning.

[0062] In one embodiment, if the noise score inference model (110) is implemented based on a non-parametric model such as KNN, the update module (244) can improve one or more policies of the noise score inference model (110) based on the evaluation of the evaluation module (242). For example, if the evaluation module (242) assigns a score greater than a predefined threshold to the output of the neural network model (120) or assigns a positive label (e.g., a label corresponding to 'good output') to the output of the neural network model (102), the update module (244) can adjust one or more policies of the noise score inference model (110) so that the noise score inference model (110) assigns a higher score (or a label corresponding to 'good noise') to the noise.

[0063] FIG. 5 illustrates an exemplary method for generating an output from a neural network model (120) based on noise generated using a noise distribution model (212) according to one embodiment of the present disclosure.

[0064] Referring to FIG. 5, the noise generation module (210) of FIG. 2 can sample noise from a noise distribution model (212) to be input to a neural network model (120). The noise generation module (210) can receive a condition vector and generate noise from the noise distribution model (212) based on the condition vector. The noise distribution model (212) can be implemented based on a parametric model or a non-parametric model. Unlike the embodiment shown in FIG. 3, in the embodiment of FIG. 5, the noise generated from the noise distribution model (212) can be provided directly to the inference module (230) without filtering.

[0065] In one embodiment, the noise distribution model (212) may be implemented based on a neural network model. For example, the noise distribution model (212) may be a neural network model having one or more weights that are pre-trained to generate noise suitable for a given condition vector. The noise generation module (210) may execute the noise distribution model (212) by inputting a condition vector into the noise distribution model (212). The noise distribution model (212) may receive the condition vector and output noise.

[0066] In one embodiment, the noise distribution model (212) may be implemented based on a Gaussian process. For example, the noise distribution model (212) may include an average function and one or more kernel functions. The noise generation module (210) may use the average function and one or more kernel functions included in the noise distribution model (212) to extract one or more noises from a multivariate Gaussian distribution based on a condition vector.

[0067] In one embodiment, the noise distribution model (212) may be implemented based on KDE. For example, the noise distribution model (212) may be obtained by estimating the probability density function of the noise distribution based on a plurality of condition vectors and a plurality of learning noises. The plurality of learning noises for the noise distribution model (212) may be noises that received a good evaluation when used to generate an output by the neural network model (120) based on the corresponding condition vector. To generate noise, the noise generation module (210) may extract sample noise based on the condition vector from the probability density function of the noise distribution model (212).

[0068] The inference module (230) can execute the neural network model (120). The inference module (230) can input a condition vector and noise provided from the noise generation module (210) into the neural network model (120). Based on the condition vector and the noise, the neural network model (120) can output a result that satisfies the requirements indicated by the condition vector. The output of the neural network model (120) can be provided to a user (50) of the electronic device (200).

[0069] By using a noise distribution model (212), instead of one of a plurality of noises predefined to the neural network model (120) being input, noise corresponding to a given condition vector can be input to the neural network model (120). In response to a given condition vector, the noise generation module (210) can sample one or more noises from a pre-trained noise distribution model (212). Thus, the noise distribution model (212) can ensure diversity of noises and further ensure diversity of the neural network model (120). Additionally, the noise distribution model (212) can be trained based on noises evaluated as good noise. Thus, the noise distribution model (212) can prevent the neural network model (120) from generating an inappropriate output.

[0070] Additionally or alternatively, the output generated by the neural network model (120) may be transmitted to the model update module (240). The model update module (240) may receive the condition vector and the output of the neural network model (120). The model update module (240) may update the noise distribution model (212) based on the condition vector and the output of the neural network model (120). The model update module (240) may include an evaluation module (242) and an update module (244). The evaluation module (242) may evaluate the output of the neural network model (120) based on the condition vector. The evaluation module (242) may include a reward model (402) and / or a feedback request module (404). The reward model (402) and the feedback request module (404) may quantitatively evaluate the output of the neural network model (120) as described above with reference to FIG. 4. The update module (244) can update the noise generation module (210) based on a quantitative evaluation of the output of the neural network model (120). Based on the evaluation of the reward model (402), the update module (244) can adjust one or more parameters of the noise distribution model (212) and / or one or more policies of the noise score inference model (110).

[0071] For example, the update module (244) can update the noise distribution model (212) based on an evaluation of the output obtained by the evaluation module (242). In one embodiment, the update module (244) can adjust one or more parameters or policies of the noise distribution model (212) based on the evaluation of the evaluation module (242). For example, if the evaluation module (242) assigns a score greater than a predefined threshold to the output of the neural network model (120) or assigns a positive label (e.g., a label corresponding to 'good output') to the output of the neural network model (102), the update module (244) can label the noise generated by the noise distribution model (212) as 'good noise' and update the noise distribution model (212) based on the labeled noise. In one embodiment, the update module (244) can update one or more parameters of the noise distribution model (212) using reinforcement learning.

[0072] In one embodiment, to implement the noise distribution model (212), test noise may be obtained from the noise distribution model. A test output may be obtained by inputting the test noise and the learning condition vector into the neural network model (120). An evaluation of the test output may be performed. Based on the evaluation, the noise distribution model (212) may be updated. Additionally or alternatively, performing an evaluation of the test output may include inputting the test output into a reward model (402) for inferring a score for the output of the neural network model (120) or requesting an evaluation of the test output from the user via the feedback request module (404).

[0073] In the embodiment illustrated in FIG. 5, a noise distribution model (212) for generating noise to be input to the neural network model (120) may be pre-set. The noise distribution model (212) may be pre-trained based on a number of learning conditions and learning noises. For example, the noise distribution model (212) may be obtained by modeling the distribution of noises labeled as suitable for a single learning condition. Accordingly, the noise distribution model (212) may be pre-trained to generate noise optimized for each of the various condition vectors. Implementing the noise distribution model (212) may correspond to limiting the noise distribution for each condition. The noise distribution model (212) may be pre-trained to sample data points (e.g., noise) in the optimal latent space for a given condition vector. Consequently, the quality of the output generated by the neural network model (120) may be improved.

[0074] FIG. 6 illustrates an exemplary flowchart of a method (600) for acquiring noise to be input to a neural network model based on a condition vector according to one embodiment of the present disclosure.

[0075] Referring to FIG. 6, the method (600) may include steps (602, 604, 606, 608, 610, 612). In one embodiment, the method (600) may be performed by an electronic device (200). However, the present disclosure is not limited thereto, and the steps (602, 604, 606, 608, 610, 612) may be performed individually or in combination by any electronic device. A method according to one embodiment of the present disclosure is not limited to that shown in FIG. 6, and any one of the steps shown in FIG. 6 may be omitted, or additional steps not shown in FIG. 6 may be included. In some embodiments, the order of at least some of the steps (602, 604, 606, 608, 610, 612) may be changed.

[0076] In step (602), the electronic device (200) can obtain a condition vector for the neural network model (120). In step (604), the electronic device (200) can generate noise to be input into the neural network model (120). In step (606), the electronic device (200) can input the noise and condition vector into the noise score inference model (110). In step (608), the electronic device (200) can identify whether the score output from the noise score inference model (110) is greater than a threshold. In step (610), the electronic device (200) can input the noise and condition vector into the neural network model (120) based on identifying that the score is greater than the threshold. In step (612), the electronic device (200) can regenerate the noise based on identifying that the score is not greater than the threshold.

[0077] Additionally or alternatively, to generate noise, the electronic device (200) may acquire a noise distribution model (212) representing the distribution of noise. The electronic device (200) may acquire noise by performing sampling based on a given condition vector from the noise distribution model. Additionally or alternatively, the noise distribution model (212) may be implemented based on any one or a combination of various algorithms such as a neural network model that outputs noise, a Gaussian process, KDE, SVM, and / or KNN.

[0078] Additionally or alternatively, the method (600) may further include the steps of obtaining an output corresponding to noise and condition vectors from a neural network model (120), performing an evaluation of the output, and updating a noise distribution model (212) based on the evaluation. Additionally or alternatively, the step of performing an evaluation of the output may include at least one of the steps of inputting the output to a reward model (402) for inferring a score for the output of the neural network model (120); or requesting an evaluation of the output from a user.

[0079] Additionally or alternatively, the noise score inference model (110) may be based on a neural network model that outputs noise, a Gaussian process, a KDE, an SVM, or a KNN. Additionally or alternatively, the method (600) may further include the steps of obtaining an output based on noise and condition vectors from a neural network model (120), providing the output to a user, obtaining an evaluation of the output from the user, and updating the noise score inference model (110) based on the obtained evaluation.

[0080] Additionally or alternatively, the noise score inference model (110) may be learned by obtaining test noise from a noise distribution model (212), obtaining a test output by inputting the test noise and a learning condition vector into a neural network model (120), performing a test evaluation on the test output, inferring a test score for the test output using the noise score inference model (110), and updating the noise score inference model (110) based on the test evaluation and the test score. Additionally or alternatively, performing a test evaluation on the test output may include inputting the test output into a reward model for inferring a score for the output of the neural network model (120) or requesting an evaluation of the test output from a user.

[0081] FIG. 7 illustrates an exemplary block diagram of an electronic device (700) according to one embodiment of the present disclosure.

[0082] Referring to FIG. 7, the electronic device (700) may include a processor (710), memory (720), a display device (730), and a communication device (740). FIG. 7 illustrates only essential components for explaining the function and / or operation of the electronic device (700), and the components included in the electronic device (700) are not limited to those illustrated in FIG. 7. The configuration of the electronic device (700) illustrated in FIG. 7 is merely an example, and the configuration illustrated in FIG. 7 does not limit the electronic device for performing an embodiment of the present disclosure. In one embodiment, one or more of the configurations illustrated in FIG. 7 may be deleted or modified, or a configuration not illustrated in FIG. 7 may be added to the electronic device (700).

[0083] In one embodiment, the electronic device (700) may be a portable device or a mobile device, and in these embodiments, the electronic device (700) may further include a battery that supplies driving power to a processor (710), a memory (720), a display device (730), and a communication device (740).

[0084] The processor (710) can execute one or more instructions of a program stored in memory (720). For example, the processor (710) can execute a noise score inference model (110) and a neural network model (120) stored in memory (720). The processor (710) may be composed of hardware components that perform arithmetic, logic, and input / output operations. Although the processor (710) is depicted as a single element in FIG. 7, it is not limited thereto. In one embodiment of the present disclosure, the processor (710) may be composed of one or more elements.

[0085] The processor (710) may include various processing circuits and / or multiple processors. For example, the term "processor" as used in this disclosure, including in the claims, may include at least one processor and, additionally or alternatively, various processing circuits. One or more processors may be configured to perform the various functions described in this disclosure in a distributed manner, individually and / or collectively. As used herein, "processor," "at least one processor," and "one or more processors" may be configured to perform various functions. However, these terms cover, without limitation, situations where one processor performs some of the functions and other processor(s) perform other parts of the functions, and situations where a single processor can perform all functions. Additionally, at least one processor may include a combination of processors performing various functions of the disclosed functions in a distributed manner. At least one processor may execute program instructions individually or collectively to achieve or perform various functions.

[0086] The processor (710) may be implemented as a general-purpose processor such as a CPU (Central Processing Unit), AP (Application Processor), DSP (Digital Signal Processor), a graphics-dedicated processor such as a GPU (Graphic Processing Unit) or VPU (Vision Processing Unit), or an artificial intelligence-dedicated processor such as an NPU (Neural Processing Unit). The processor (710) may be controlled to process input data according to predefined operation rules or an artificial intelligence model. Alternatively, if the processor (710) is an artificial intelligence-dedicated processor, the artificial intelligence-dedicated processor may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0087] The memory (720) can store instructions, data structures, and program code that can be read by the processor (710). For example, the memory (720) can store a noise score inference model (110), a neural network model (120), and a noise distribution model (212) that are readable (or executable) by the processor (710). In one embodiment, the memory (720) can store instructions that can cause the electronic device (700) to perform at least some of the operations of the electronic device (200) of FIG. 2 described above with reference to FIG. 1 through 6 by being executed individually or in combination by the processor (710). For example, the processor (710) can perform at least some of the operations described in FIG. 1 through 6 by executing one or more instructions or codes stored in the memory (720).

[0088] The memory (720) may include flash memory type, hard disk type, multimedia card micro type, and card type memory, and may include non-volatile memory including at least one of ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk, and / or volatile memory such as DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory).

[0089] In one embodiment, the memory (720) may store one or more instructions and / or program codes that cause the electronic device (700) to generate noise, perform score-based filtering based on a given condition vector for the noise, and execute a neural network model (120) using the filtered noise and the given condition vector. For example, the memory (720) may store instructions and / or program codes for implementing at least some of the functions of the noise generation module (210), noise filtering module (220), inference module (230), or model update module (240) of FIG. 2. Meanwhile, the elements stored in the memory (720) described above are for convenience of explanation and are not necessarily limited thereto.

[0090] The display device (730) can output an image signal to the screen of the electronic device (700) under the control of the processor (710). For example, the display device (730) can output an image or video to display the result generated from the neural network model (120) by the electronic device (700).

[0091] In one embodiment, the display device (730) may include a touch panel. The touch panel may include one or more touch sensors that detect touch input. In one embodiment, a condition vector may be input from a user (70) through the touch panel. In one embodiment, feedback on an output generated using a neural network model (120) may be input through the touch panel.

[0092] The communication device (740) can perform data communication with other external devices under the control of the processor (710). In one embodiment, the communication device (740) may include communication circuit(s) capable of performing data communication between an electronic device (700) and another electronic device using at least one of a data communication method including wired LAN (Local Area Network, LAN), wireless LAN (Wi-Fi), Bluetooth, ZigBee, WFD (Wi-Fi Direct), infrared communication (IrDA, infrared Data Association), BLE (Bluetooth Low Energy), NFC (Near Field Communication), Wibro (Wireless Broadband Internet, Wibro), WiMAX (World Interoperability for Microwave Access, WiMAX), SWAP (Shared Wireless Access Protocol), WiGig (Wireless Gigabit Alliances, WiGig), and RF communication.

[0093] In the embodiment of FIG. 7, the electronic device (700) may include a display device (730). The electronic device (700) may display the output of a neural network model (120) corresponding to generated noise and a given condition vector on the display device (730). Accordingly, the output of the neural network model (120) may be provided to a user (70).

[0094] FIG. 8 illustrates an exemplary flowchart of a method (800) for providing an output from a neural network model (120) to a user in response to a condition vector input by a user (70) according to one embodiment of the present disclosure.

[0095] Referring to FIG. 8, the method (800) may include steps (802, 804, 806, 808, 810, 812, 814, 816). In one embodiment, the method (800) may be performed by an electronic device (700) and a user (70). However, the present disclosure is not limited thereto, and the steps (802, 804, 806, 808, 810, 812, 814, 816) may be performed individually or in combination by any electronic device. A method according to one embodiment of the present disclosure is not limited to that shown in FIG. 8, any one of the steps shown in FIG. 8 may be omitted, and additional steps not shown in FIG. 8 may be included. In some embodiments, the order of at least some of the steps (802, 804, 806, 808, 810, 812, 814, 816) may be changed.

[0096] In step (802), the user (70) can input a condition vector into the electronic device (700). In step (804), the electronic device (700) can generate noise to be input into the neural network model (120), for example, using a noise distribution model (212) and / or a noise score inference model (110). In step (806), the electronic device (700) can generate an output corresponding to the condition vector from the neural network model (120) based on the noise and the condition vector. In step (808), the electronic device (700) can provide the generated output to the user (70). In step (810), the electronic device (700) can request the user (70) to evaluate the output generated from the neural network model (120). In response to the request from the electronic device (700), in step (812), the user (70) can perform an evaluation of the output. In step (814), the user (70) may provide the evaluation performed in step (812) to the electronic device (700). In step (816), based on the evaluation, the electronic device (700) may update at least one of the noise distribution model (212) or the noise score inference model (110).

[0097] FIG. 9 illustrates, in accordance with one embodiment of the present disclosure, a user interface (UI) displayed on a display device (730) of FIG. 7.

[0098] Referring to FIG. 9, the electronic device (700) of FIG. 7 may display a first graphic UI (Graphic UI, GUI) (910) and a second GUI (920) on an area of ​​a display device (730). The first GUI (910) may be a GUI for obtaining a condition vector for a neural network model (120). The second GUI (920) may be a GUI for providing the output of the neural network model (120) to a user (70).

[0099] The first GUI (910) may include an input request UI (912) for requesting prompt input from a user and an input display UI (914) for displaying prompts entered by the user. The input request UI (912) may include a request for the user (70) to input a condition vector for the neural network model (120). For example, the input request UI (912) may include text including a request for the user (70) to input a condition vector for the neural network model (120). The input display UI (914) may include prompts entered by the user (70) in response to the request of the input request UI (912). For example, the input display UI (914) may display user-entered prompts such as 'cats and dogs'.

[0100] Based on a user-input prompt entered through the first GUI (910), the electronic device (700) can generate noise and execute a neural network model (120) based on the generated noise. For example, the electronic device (700) can obtain a condition vector representing the user's (70) requirements by preprocessing the prompt entered by the user (70). The electronic device (700) can obtain one or more noises from a noise distribution model (214). The electronic device (700) can generate one or more outputs by inputting one or more noises into the neural network model (120). Additionally or alternatively, the electronic device (700) can infer the scores of the obtained noises using a noise score inference model (110). The electronic device (700) can generate one or more outputs by inputting noise(s) whose inferred scores are greater than a threshold into the neural network model (120).

[0101] The second GUI (920) may include an evaluation request UI (922), an output display UI (924), an output selection UI (926), and an input display UI (914). The evaluation request UI (922) may include text for requesting the user (70) to evaluate each output. For example, the evaluation request UI (922) may include text for requesting the user (70) to select the most preferred output among the outputs from the neural network model (120). In one embodiment, the evaluation request UI (922) may include text for requesting the user (70) to evaluate each output of the neural network model (120) (e.g., text requesting to enter a score for each output or text requesting to enter a comment for each output).

[0102] The output display UI (924) may include at least one of the results output from the neural network model (120). For example, through the output display UI (924), the output of the neural network model (120) based on 'Noise A' among one or more noises acquired by the electronic device (200) may be displayed on the display device (730). The output of the neural network model (120) displayed on the display device (730) through the output display UI (924) may be determined by the input of the user (70) to the output selection UI (926).

[0103] The output selection UI (926) may include one or more buttons corresponding to one or more outputs generated by the neural network model (120). In the embodiment illustrated in FIG. 9, button 'A' of the output selection UI (926) may correspond to the output of the neural network model (120) based on 'noise A'. Button 'B' of the output selection UI (926) may correspond to the output of the neural network model (120) based on 'noise B'. Button 'C' of the output selection UI (926) may correspond to the output of the neural network model (120) based on 'noise C'. In response to input from the user (70) to any one of the buttons included in the output selection UI (926), the output displayed through the output display UI (924) may change. For example, in response to input from the user (70) to button 'A' of the output selection UI (926), the output of the neural network model (120) based on 'noise A' can be displayed through the output display UI (924).

[0104] In response to an evaluation request via an evaluation request UI (922), the electronic device (700) may obtain at least one evaluation from the user (70) for at least one of the outputs of the neural network model (120). Based on at least one evaluation obtained from the user (70), the electronic device (700) may update at least one of the noise score inference model (110) or the noise distribution model (212). For example, the electronic device (700) may update at least one of the noise score inference model (110) or the noise distribution model (212) in a manner similar to the manner described above with reference to FIG. 4 and / or FIG. 5.

[0105] Through the first GUI (910) and the second GUI (920), the electronic device (700) can obtain evaluations from the user (70) regarding the outputs of the neural network model (120). Based on the evaluations from the user (70), the electronic device (700) can update at least one of the noise score inference model (110) or the noise distribution model (212). Accordingly, the noise score inference model (110) and / or the noise distribution model (212) can be customized based on the user's (70) preferences, and consequently, the quality of the output of the neural network model (120) can be improved.

[0106] FIG. 10 illustrates an exemplary block diagram of an electronic device (1000) according to one embodiment of the present disclosure.

[0107] Referring to FIG. 10, the electronic device (1000) may include a processor (1010), memory (1020), and a communication device (1030). FIG. 10 illustrates only essential components for explaining the function and / or operation of the electronic device (1000), and the components included in the electronic device (1000) are not limited to those illustrated in FIG. 10. The configuration of the electronic device (1000) illustrated in FIG. 10 is merely an example, and the configuration illustrated in FIG. 10 does not limit the electronic device for performing an embodiment of the present disclosure. In one embodiment, one or more of the configurations illustrated in FIG. 10 may be deleted or changed, or a configuration not illustrated in FIG. 10 may be added to the electronic device (1000).

[0108] The processor (1010) can execute one or more instructions of a program stored in memory (1020). For example, the processor (1010) can execute a noise score inference model (110) and a neural network model (120) stored in memory (1020). The processor (1010) may be composed of hardware components that perform arithmetic, logic, and input / output operations. Although the processor (1010) is depicted as a single element in FIG. 10, it is not limited thereto. In one embodiment of the present disclosure, the processor (1010) may be composed of one or more elements.

[0109] The processor (1010) may include various processing circuits and / or multiple processors. For example, the term "processor" as used in this disclosure, including in the claims, may include at least one processor and, additionally or alternatively, may include various processing circuits. One or more processors may be configured to perform the various functions described in this disclosure in a distributed manner, individually and / or in combination. As used herein, "processor," "at least one processor," and "one or more processors" may be configured to perform various functions. However, these terms cover, without limitation, situations where one processor performs some of the functions and other processor(s) perform other parts of the functions, and situations where a single processor can perform all functions. Additionally, at least one processor may include a combination of processors performing various functions of the disclosed functions in a distributed manner. At least one processor may execute program instructions individually or in combination to achieve or perform various functions.

[0110] The processor (1010) may be implemented as a general-purpose processor such as a CPU, AP, DSP, etc., a graphics-dedicated processor such as a GPU, VPU, or an artificial intelligence-dedicated processor such as an NPU. The processor (1010) may be controlled to process input data according to predefined operation rules or an artificial intelligence model. Alternatively, if the processor (1010) is an artificial intelligence-dedicated processor, the artificial intelligence-dedicated processor may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0111] The memory (1020) may store instructions, data structures, and program code that can be read by the processor (1010). For example, the memory (1020) may store a noise score inference model (110), a neural network model (120), and a noise distribution model (212) that are readable (or executable) by the processor (1010). In one embodiment, the memory (1020) may store instructions that can cause the electronic device (1000) to perform at least some of the operations of the electronic device (200) of FIG. 2 described above with reference to FIG. 1 through 6 by being executed individually or in combination by the processor (1010). For example, the processor (1010) may perform at least some of the operations described in FIG. 1 through 6 by executing one or more instructions or codes stored in the memory (1020).

[0112] The memory (1020) may include a flash memory type, a hard disk type, a multimedia card micro type, a card type memory, and may include a non-volatile memory including at least one of a ROM, EEPROM, PROM, magnetic memory, a magnetic disk, and an optical disk, and / or a volatile memory such as a DRAM or SRAM.

[0113] In one embodiment, the memory (1020) may store one or more instructions and / or program codes that cause the electronic device (1000) to generate noise, perform score-based filtering based on a given condition vector for the noise, and execute a neural network model (120) using the filtered noise and the given condition vector. For example, the memory (1020) may store instructions and / or program codes for implementing at least some of the functions of the noise generation module (210), noise filtering module (220), inference module (230), or model update module (240) of FIG. 2. Meanwhile, the elements stored in the memory (1020) described above are for convenience of explanation and are not necessarily limited thereto.

[0114] A communication device (1030) can perform data communication with an external device (12) under the control of a processor (1010). For example, the communication device (1030) can receive a condition vector from an external device (12) that communicates with a user (14). The external device (12) may be a terminal or equipment of the user (14). In one embodiment, the communication device (1030) may include communication circuit(s) capable of performing data communication between an electronic device (1000) and another electronic device using at least one of a data communication method including wired LAN, wireless LAN, Wi-Fi, Bluetooth, Zigbee, WFD, infrared communication, BLE, NFC, WiBro, WiMAX, SWAP, WiGig, and RF communication.

[0115] In the embodiment of FIG. 10, the electronic device (1000) may include a communication device (1030) that receives a condition vector from an external device (12). The electronic device (1000) may transmit the output of a neural network model (120) corresponding to the generated noise and the given condition vector to the external device (12) through the communication device (1030). Accordingly, the output of the neural network model (120) may be provided to the user (14) by the external device (12).

[0116] FIG. 11 illustrates an exemplary flowchart of a method for providing an output from a neural network model to a user in response to a condition vector input by a user, according to one embodiment of the present disclosure.

[0117] Referring to FIG. 11, the method (1100) may include steps (1102, 1104, 1106, 11011, 1110, 1112, 1114, 1116, 1118, 1120, 1122). In one embodiment, the method (1100) may be performed by an electronic device (1000), an external device (12), and a user (15). However, the present disclosure is not limited thereto, and the steps (1102, 1104, 1106, 11011, 1110, 1112, 1114, 1116, 1118, 1120, 1122) may be performed individually or in combination by any electronic device. A method according to one embodiment of the present disclosure is not limited to that illustrated in FIG. 11, and any one of the steps illustrated in FIG. 11 may be omitted, or additional steps not illustrated in FIG. 11 may be included. In some embodiments, the order of at least some of the steps (1102, 1104, 1106, 11011, 1110, 1112, 1114, 1116, 1118, 1120, 1122) may be changed.

[0118] In step (1102), the user (14) can input a condition vector to the external device (12). In step (1104), the external device (12) can transmit the condition vector input from the user (14) to the electronic device (1000). In step (1106), the electronic device (1000) can generate noise to be input to the neural network model (120), for example, using a noise distribution model (212) and / or a noise score inference model (110). In step (1108), the electronic device (1000) can generate an output corresponding to the condition vector from the neural network model (120) based on the noise and the condition vector. In step (1110), the electronic device (1000) can transmit the output generated from the neural network model (120) to the external device (12). In step (1112), the electronic device (1000) may request an evaluation of the output generated from the neural network model (120) from the external device (12). In response to the request from the electronic device (1000), in step (1114), the external device (12) may request an evaluation of the output along with providing the generated output to the user (14). In step (1116), the user (14) may perform an evaluation of the output, for example, in a manner similar to the embodiment of FIG. 9. In step (1118), the user (14) may provide the evaluation performed in step (1116) to the external device (12). In step (1120), the external device (12) may provide the evaluation of the user (14) to the electronic device (1000). In step (1122), based on the evaluation, the electronic device (1000) can update at least one of the noise distribution model (212) or the noise score inference model (110).

[0119] A device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-transitory storage medium' simply means that it is a tangible device and does not contain a signal (e.g., electromagnetic waves), and the term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily. For example, a 'non-transitory storage medium' may include a buffer in which data is stored temporarily.

[0120] 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 or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., 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.

[0121] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various changes and modifications from the description above. For example, appropriate results can be achieved even if the described techniques are performed in a different order than described, and / or components such as the described computer system or module are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

Claims

A step of obtaining a condition vector for a neural network model (120); A step of generating noise to be input into the above neural network model 120); A step of inputting the noise and the condition vector into a noise score inference model (110); A step of identifying whether the score output from the above noise score inference model (110) is greater than a threshold; A step of inputting the noise and the condition vector into the neural network model (120) based on identifying that the score is greater than the threshold; and A method comprising the step of regenerating noise based on identifying that the above score is not greater than the above threshold. In paragraph 1, The step of generating the noise to be input into the neural network model (120) is: A step of obtaining a noise distribution model (212) representing the distribution of the above noise; and A method comprising the step of obtaining the noise by performing sampling based on the condition vector from the noise distribution model (212). In paragraph 2, The above noise distribution model (212) is a method based on a neural network model that outputs noise, a Gaussian process, Kernel Density Estimation (KDE), Support Vector Machine (SVM), or K-Nearest Neighbor (KNN). In paragraph 2 or 3, A step of obtaining an output corresponding to the noise and the condition vector from the above neural network model (120); A step of performing an evaluation of the above output; and A method further comprising the step of updating the noise distribution model (212) based on the above evaluation. In paragraph 4, The step of performing an evaluation of the above output is: A step of inputting the output to a reward model for inferring a score for the output of the above neural network model (120); or A method comprising at least one step of requesting a user to evaluate the above output. In any one of paragraphs 1 through 5, The above noise score inference model (110) is a method based on a neural network model that outputs noise, a Gaussian process, KDE, SVM, or KNN. In any one of paragraphs 1 through 6, A step of obtaining an output based on the noise and the condition vector from the above neural network model (120); A step of providing the above output to a user; A step of obtaining an evaluation of the output from the above user; and A method further comprising the step of updating the noise score inference model (110) based on the above-mentioned evaluation. In any one of paragraphs 1 through 7, The above noise score inference model (110) is, Obtaining test noise from the above noise distribution model (212); Obtaining a test output by inputting the above test noise and learning condition vector into the above neural network model (120); Performing a test evaluation of the above test output; Inferring a test score for the test output using the above noise score inference model (110); and A method learned by updating the noise score inference model (110) based on the above test evaluation and the above test score. In paragraph 8, A method of performing a test evaluation of the above test output, comprising inputting the above test output into a reward model (402) for inferring a score for the output of the above neural network model (120) or requesting an evaluation of the above test output from a user. A computer-readable recording medium having a program recorded thereon for performing any one of the methods of paragraphs 1 through 9 on a computer. As an electronic device (700, 1000), At least one processor (710, 1010) including a processing circuit; and It includes one or more storage media and a memory (720, 1020) that stores one or more instructions, and When the above one or more instructions are executed individually or in combination by the above at least one processor, the electronic device: Obtain a condition vector for the neural network model; Generate noise to be input into the above neural network model; Input the above noise and the above condition vector into a noise score inference model; Identify whether the score output from the above noise score inference model is greater than a threshold; Based on identifying that the above score is greater than the above threshold, the noise and the condition vector are input into the neural network model; and An electronic device that causes noise to be regenerated based on identifying that the above score is not greater than the above threshold. In Paragraph 11, The above electronic device further includes a display device (730), and An electronic device that, when the above one or more instructions are executed individually or in combination by the above at least one processor, causes the electronic device to additionally display the output of the neural network model corresponding to the noise and the condition vector on the display device (730). In Paragraph 11, The electronic device further includes a communication device (1030) that receives the condition vector from an external device, and An electronic device that, when the above one or more instructions are executed individually or in combination by the above at least one processor, causes the electronic device to additionally transmit the output of the neural network model corresponding to the noise and the condition vector to the external device through the communication device (1030). In any one of paragraphs 11 through 13, When the above one or more instructions are executed individually or in combination by the above at least one processor, the electronic device additionally: Obtaining a noise distribution model representing the distribution of the above noise; and An electronic device that obtains the noise by performing sampling based on the condition vector from the noise distribution model. In any one of paragraphs 11 through 14, When the above one or more instructions are executed individually or in combination by the above at least one processor, the electronic device additionally: Obtaining an output based on the noise and the condition vector from the above neural network model; Providing the above output to the user; Obtaining an evaluation of the output from the above user; and An electronic device that updates the noise score inference model based on the above-mentioned evaluation.

Images