Information processing system, information processing method, and storage medium
The information processing system addresses the lack of element-specific evaluation in TDA by using Concept-TRAK to calculate the influence of training data on elements, enhancing the precision and stability of attribution analysis.
Patent Information
- Application Number
- PCT/JP2026/007023
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2026-02-26
- Publication Date
- 2026-11-12
AI Technical Summary
Conventional techniques for Training Data Attribution (TDA) in machine learning models fail to specify the elements to be evaluated, leading to insufficient evaluation of training data for specific elements such as styles and objects.
An information processing system that evaluates machine learning training data by targeting specific elements using an element evaluation function, calculating the influence of each training data point on elements like styles and objects through a framework called Concept-TRAK, which includes a robust loss function and geometrically motivated gradients to provide element-specific attribution.
Enables accurate and stable evaluation of training data influence on specific elements, allowing for precise attribution analysis and improved model performance by focusing on elements like styles and objects.
Smart Images

Figure JP2026007023_12112026_PF_FP_ABST
Abstract
Description
Information processing system, information processing method, and storage medium
[0001] This disclosure relates to an information processing system, an information processing method, and a storage medium.
[0002] Conventionally, techniques for Training Data Attribution (TDA), which estimates the influence of each data point used to train a machine learning model (also called "training data"), have been provided (for example, Non-Patent Document 1).
[0003] Xiaosen Zheng, Tianyu Pang, Chao Du, Jing Jiang, Min Lin, "Intriguing Properties of Data Attribution on Diffusion Models", [online] [Retrieved January 22, 2026], Internet <URL: https: / / arxiv.org / pdf / 2311.00500>
[0004] However, there is room for improvement in conventional techniques. For example, while conventional techniques can extend TDA to diffusion models and estimate and evaluate the overall impact of each training data point, they do not consider specifying the elements to be evaluated, such as the elements included in the training data. Therefore, there is a need to appropriately evaluate machine learning training data according to specifications.
[0005] Therefore, this disclosure proposes an information processing system, an information processing method, and a storage medium that can appropriately evaluate machine learning training data according to specifications.
[0006] To solve the above problems, one form of information processing system according to this disclosure includes a processing circuit, which receives content generated by a machine learning model, receives designation information that specifies elements included in the content, and evaluates machine learning training data with respect to the elements specified by the designation information.
[0007] This figure shows an example of information processing according to the embodiment of this disclosure. This figure shows an overview of learning and inference related to a content generation model. This figure shows an example configuration of an information processing system according to the embodiment of this disclosure. This is a flowchart of the information processing procedure according to the embodiment of this disclosure. This figure shows an example of another application example. This is a hardware configuration diagram showing an example of a computer that realizes the functions of an information processing device.
[0008] Embodiments of this disclosure will be described in detail below with reference to the drawings. Note that these embodiments do not limit the information processing system, information processing method, and storage medium described herein. Furthermore, in each of the following embodiments, the same parts are denoted by the same reference numerals to avoid redundant explanations.
[0009] This disclosure will be described in the following order of items: 1. Embodiments 1-1. Technical Overview (Background, etc.) 1-2. Overview of Information Processing According to the Embodiments of This Disclosure 1-2-1. Influence Function 1-3. Configuration of the Information Processing System According to the Embodiments 1-4. Procedure for Information Processing According to the Embodiments 1-5. Other Processing Examples (Music, etc.) 2. Others 3. Hardware Configuration
[0010] <1. Embodiments> <1-1. Technical Overview (Background, etc.)> Before describing the information processing etc. according to the embodiments of this disclosure, an overview of existing technologies related to TDA and their problems will be provided. Detailed explanations of the processing etc. of existing technologies related to TDA will be omitted as appropriate.
[0011] Techniques related to TDA are used to quantify the influence of each training data on the inference of machine learning models. Examples of such TDA techniques include D-TRAK disclosed in Non-Patent Document 1, as well as TRAK disclosed in the following document: • Sung Min Park, et al. "TRAK: Attributing Model Behavior at Scale." In Proceedings of the 40th International Conference on Machine Learning, pages 27074-27113, 2023.
[0012] However, as mentioned above, existing TDA technologies do not consider specifying the elements to be evaluated, such as the elements included in the training data. Thus, while existing TDA technologies can evaluate data at the overall level of content such as images and identify influential data, they are insufficient for evaluating data for specific elements such as styles and objects.
[0013] <1-2. Overview of Information Processing According to the Embodiment of the Disclosure> The information processing system 1 (see Figure 3), which has an information processing device 100, performs information processing as outlined in Figure 1. Figure 1 is a diagram showing an example of information processing according to the embodiment of the disclosure. Details of the configuration of the information processing system 1 will be described later, but the information processing device 100 in the information processing system 1 is an example of a computer that performs processing related to the evaluation of machine learning training data.
[0014] In Figure 1, the information processing device 100 evaluates the training data included in the training dataset used to train the content generation model M1, which is an example of a machine learning model, from the perspective of each element. For example, Figure 1 illustrates the case where the information processing device 100 evaluates the training data used to train the content generation model M1, which generates images, targeting each of the styles and objects specified as elements. Note that the content is not limited to images, but may be various types of content, but this point will be discussed later. Furthermore, the evaluation of training data as referred to here includes various processes related to the value (evaluation) of the training data, such as calculating the influence of the training data, identifying trading data, and ranking trading data.
[0015] First, the information processing device 100 generates content CT1, which is an image (step S1). In Figure 1, the information processing device 100 generates content CT1 by inputting prompt PT1 to the content generation model M1 and causing the content generation model M1 to output content CT1, which is an image. For example, prompt PT1 is instruction information requesting the generation of an image containing the string "Cat in the style of graffiti art". In this way, the information processing device 100 generates content CT1, which is an image containing a cat in the style of graffiti art, by inputting prompt PT1, which contains instruction information requesting the generation of an image containing a cat in the style of graffiti art, to the content generation model M1. The information processing device 100 may also acquire content CT1 without generating it. For example, the information processing device 100 may acquire content CT1 from a user requesting evaluation of training data. For example, the information processing device 100 may acquire content CT1 based on user input.
[0016] Here, we will explain the learning and inference processes for the content generation model M1 using Figure 2. Figure 2 is a diagram illustrating the overview of the learning and inference processes for the content generation model. For example, Figure 2 shows the training data used to learn the content generation model M1 and an image of the content generated by the content generation model M1.
[0017] As shown in Figure 2, the content generation model M1 is trained using a training dataset DS1 that includes multiple training data such as training data TD1 to TD6, each of which is an image (hereinafter also referred to as "training data TD") (corresponding to "Training" in Figure 2). Note that in Figure 2, only the six data sets TD1 to TD6 are shown, but the training dataset DS1 contains a large number of images (e.g., millions, hundreds of millions, etc.) as training data. For example, the training dataset DS1 is a dataset stored in the memory of the information processing device 100, such as the memory unit 120 (see Figure 3).
[0018] Each training data TD included in the training dataset DS1 shown in Figure 2 is an image containing a specific object, such as a cat, or an image containing a specific style, such as graffiti on a wall. For example, training data TD1 is an image containing graffiti on a wall drawn in a specific style. Also, for example, training data TD2 is an image containing a cat, which is an example of an object.
[0019] The learning of the content generation model M1 may be performed by the information processing device 100, or by a device other than the information processing device 100 (such as a learning device). When a device other than the information processing device 100 (such as a learning device) learns the content generation model M1, the information processing device 100 acquires the content generation model M1 by receiving it from the device that learned the content generation model M1.
[0020] As shown in Figure 2, the content generation model M1 is used to generate content such as content CT1 (corresponding to "Inferece" in Figure 2). For example, in Figure 1, when prompt PT1 containing the string "Cat in the style of graffiti art" is input to the content generation model M1, it outputs content CT1, which is an image containing a cat in the style of graffiti art.
[0021] The information processing device 100 then performs a process to evaluate the training data TD used to train the content generation model M1, targeting the elements contained in the content CT1. For example, the information processing device 100 performs a process to evaluate the training data TD used to train the content generation model M1, targeting the elements related to objects contained in the content CT1 (step S2). In Figure 1, the information processing device 100 performs a process to evaluate the training data TD used to train the content generation model M1, targeting the cat, which is an element related to objects contained in the content CT1. The cat element may be specified by the user, or it may be specified using an element extracted by image analysis of the content CT1.
[0022] For example, the information processing device 100 uses an influence function ("element evaluation function") to evaluate the training data TD used to train the content generation model M1, targeting the element "cat". Details of the element evaluation function will be described later, but the element evaluation function is, for example, equation (37) described later, and is a function that includes terms corresponding to elements and terms corresponding to the data to be processed.
[0023] The information processing apparatus 100 uses the element evaluation function to calculate the degree of influence of each training data TD included in the training data set DS1 when the element "cat" is targeted. Then, the information processing apparatus 100 performs higher evaluation in descending order of the degree of influence of each training data TD targeting the element "cat". For example, the information processing apparatus 100 evaluates that the larger the degree of influence of each training data TD targeting the element "cat", the greater the contribution to the generation of the element "cat" in the content generation model M1.
[0024] In FIG. 1, the information processing apparatus 100 calculates the degree of influence in descending order of training data TD3, TD100, TD75, TD6... among the training data set DS1. Therefore, as shown in the evaluation result RS1, the information processing apparatus 100 gives higher evaluation in the order of training data TD3, TD100, TD75, TD6... among the training data set DS1 when the element "cat" is targeted. For example, the information processing apparatus 100 provides the evaluation result RS1 to the user who requested the evaluation.
[0025] For example, the information processing apparatus 100 executes a process of evaluating the training data TD used for learning the content generation model M1, targeting elements related to the style included in the content CT1 (step S3). In FIG. 1, the information processing apparatus 100 executes a process of evaluating the training data TD used for learning the content generation model M1, targeting the graffiti art style which is an element related to the style included in the content CT1. Note that the specification of the graffiti art style as an element may be performed by the user, or may be a specification using an element extracted by performing image analysis on the content CT1.
[0026] For example, the information processing apparatus 100 uses the element evaluation function to execute a process of evaluating the training data TD used for learning the content generation model M1, targeting the element "graffiti art style".
[0027] The information processing device 100 uses an element evaluation function to calculate the influence of each training data TD included in the training dataset DS1 when targeting the element "graffiti art style". The information processing device 100 then gives a higher evaluation to each training data TD targeting the element "graffiti art style" in descending order of influence. For example, the information processing device 100 evaluates that the training data TD targeting the element "graffiti art style" has a greater influence, as it contributes more to the generation of the element "graffiti art style" in the content generation model M1.
[0028] In Figure 1, the information processing device 100 calculates the influence of the training data TD4, TD891, TD14, TD8, etc., in that order from the training dataset DS1. Therefore, as shown in the evaluation result RS2, the information processing device 100 gives a high evaluation to the element "graffiti art style" in that order from the training dataset DS1, in that order from the training data TD4, TD891, TD14, TD8, etc. For example, the information processing device 100 provides the evaluation result RS2 to the user who requested the evaluation.
[0029] <1-2-1. Influence function> From here, an example of an element evaluation function, which is an influence function used by the information processing device 100 in the information processing described above, will be explained. In the following explanation, existing technologies such as existing influence functions will be explained to the minimum extent necessary for explaining the element evaluation function, but detailed explanations of existing technologies will be omitted as appropriate. Also, the element evaluation function shown below is merely an example, and the information processing device 100 may use various functions as influence functions to perform the information processing described above, as long as evaluation according to the element is possible. Furthermore, the information processing device 100 may perform information processing using various models, not limited to influence functions, as long as evaluation according to the element is possible, but this point will be discussed later.
[0030] The purpose of data attribution methods such as TDA is to quantify the performance of a model using performance metrics and objective functions (such as test loss). An example of this is the training sample x, which is an example of trading data (also called a "training sample").i The goal is to evaluate (estimate) the influence (contribution) of [something].
[0031] For example, the influence function disclosed in the following document uses gradient-based estimation to train sample x i This efficiently approximates the effect of removing the black-box function. For the parameters shown in equation (1) and the training loss model shown in equation (2), the influence function is defined as shown in equation (3). • Pang Wei Koh and Percy Liang, "Understanding black-box predictions via influence functions." In International conference on machine learning, pages 1885-1894. PMLR, 2017.
[0032]
[0033]
[0034]
[0035] Here, equation (4) is given by training sample x0 i This represents the gradient of the loss with respect to the parameter θ. Equation (5) is the gradient of the utility loss (V in calligraphic form in the equation), and equation (6) is the Hessian matrix of the training loss calculated over the entire training dataset shown in equation (7).
[0036]
[0037]
[0038]
[0039]
[0040] However, calculating the influence function described above is computationally difficult because each attribution query requires not only a costly Hessian calculation but also the recalculation of the training gradient across the entire training set (see the following reference): • Sang Keun Choe et al. "What is your data worth to gpt? llm-scale data valuation with influence functions." arXiv preprint arXiv:2405.13954, 2024.
[0041] To address this problem, TRAK proposes projecting the gradient into a lower-dimensional space using a random projection matrix (see equations (8) and (9)).
[0042]
[0043]
[0044] The elements shown in equation (10) within equation (9) are projected Hessians. This enables efficient storage and reuse of gradients for multiple attribution queries.
[0045]
[0046] Regarding data attribution in diffusion models, previous research on diffusion models, such as the following literature, primarily focuses on evaluating the attribution of the entire image, and typically uses the same objective function for both training and practical losses. • Tong Xie et al. "Data attribution for diffusion models: Timestep-induced bias in influence estimation." Transactions on Machine Learning Research, 2024. • Xiaosen Zheng et al. "Intriguing properties of data attribution on diffusion models." In The Twelfth International Conference on Learning Representations, 2024.
[0047] The above research reveals that attribution evaluation performance is extremely sensitive to the choice of loss function. In particular, the standard DSM (Denoising Score Matching) loss is a noise term ε and an input perturbation x t Because probabilism is introduced through both methods, reliable estimation results in noisy gradient estimations that require large-scale averaging, making it difficult to appropriately estimate attribution evaluation. Therefore, D-TRAK, as described above, uses the squared l² norm shown in equation (11), while DAS, disclosed in the following literature, uses the squared l¹ norm shown in equation (12) to calculate the influence score, achieving improved stability and accuracy. • Jinxu Lin et al. "Diffusion attribution score: Evaluating training data influence in diffusion model." In The Thirteenth International Conference on Learning Representations, 2025.
[0048]
[0049]
[0050] The existing techniques described above indicate that selection of a robust loss function for gradient calculation is essential for highly reliable attribution analysis in diffusion models. Accordingly, the present method extends the existing techniques to element-level settings such as concepts and entities. In such a setting of the present method, a loss function specially designed to capture element-specific influences such as concepts and entities as shown below is used.
[0051] For example, the present method introduces a framework (also referred to as "Concept-TRAK") that quantifies the influence (contribution degree) of individual training samples on specific elements (concepts, entities, etc.) learned by the diffusion model. In the present method, Concept-TRAK targets evaluation (analysis) of the influence of training data on the model's ability to represent specific elements.
[0052] Element-level attribution is defined as evaluating (measuring) the influence of training sample x0 i on the model's generation capability for element c target , and is quantified as shown in Equation (13) through the element occurrence rate.
[0053]
[0054] Here, p(c target |x0) represents the probability that element c target is present in image x0, and p sample (·|c) represents the sample distribution used for generation. For example, p sample represents the generation distribution of the model (for example, conditional sampling using CFG (Classifier-Free Guidance Scale)), and corresponds to global attribution that measures element probabilities in all generations. Further, p as shown in Equation (14) sample measures the influence of the training data on the specific manifestation of the element in a specific generated image x0 test . For example, in order to evaluate (estimate) the influence (contribution degree) of the training sample on the element probability, an influence function framework as shown in Equation (15) is adopted.
[0055]
[0056]
[0057] Here, the term shown in equation (16) of equation (15) corresponds to the model performance for a specific element generation (e.g., utility loss), and the term shown in equation (17) of equation (15) corresponds to the contribution of the training samples.
[0058]
[0059]
[0060] For example, p θ (c target This involves using ) as the utility loss and a standard DSM loss as the training loss. However, as mentioned above, previous studies have shown that attribution performance is very sensitive to the design of the loss function. Therefore, instead of a general-purpose denoising signal, we design a loss function that is robust and provides element-specific gradients. For example, based on the hypothesis that the directionality of meaningful elements corresponds to tangent vectors in the latent space of the diffusion model, we design an element-aware loss function through geometric motivation.
[0061] For example, the latent variable x in a diffusion model. t These exist on low-dimensional manifolds, as disclosed in the following literature: • Hyungjin Chung et al. "Improving diffusion models for inverse problems using manifold constraints." Advances in Neural Information Processing Systems, 35:25683-25696, 2022. • Luping Liu et al. "Pseudo numerical methods for diffusion models on manifolds." In International Conference on Learning Representations, 2022.
[0062] Previous studies have identified semantically rich structures within the tangent spaces of manifolds, enabling concept-based editing methods (see the following reference): • Yong-Hyun Park et al. "Understanding the latent space of diffusion models through the lens of riemannian geometry." Advances in Neural Information Processing Systems, 36:24129-24142, 2023.
[0063] Therefore, in this method, the directionality associated with elements is manipulated in tangent space rather than with a standard denoising objective function. In this method, as shown in equation (18), the reward gradient for reward optimization functions as an element-specific guidance direction that points to a region in tangent space that reinforces the element. In this method, considering the goal of reward optimization, equation (19) is defined (see references below).・Natasha Jaques et al. "Sequence tutor: Conservative fine-tuning of sequence generation models with KL-control." In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1645-1654. PMLR, 06-11 Aug 2017. ・Rafael Rafailov et al. "Direct preference optimization: Your language model is secretly a reward model." Advances in Neural Information Processing Systems, 36:53728-53741, 2023.
[0064]
[0065]
[0066] In equation (19), R(x0) corresponds to the reward function, and p θ (•|c) corresponds to the target model for reward optimization, and β is used to control the strength of regularization. Sampling distribution p sample (•|c) can vary, but for theoretical clarity, p0(•|c) is used in the derivation. In this method, equation (20) is defined as the optimal solution (see Rafael's literature above and the literature below). • Tomasz Korbak et al. "On reinforcement learning and distribution matching for fine-tuning language models with no catastrophic forgetting." URL<https: / / arxiv.org / pdf / 2206.00761>
[0067]
[0068] In diffusion models, the reward R(x) for noisy latent variables is calculated as follows: t This can be extended to intermediate time steps by defining (see the following reference): • Bram Wallace et al. "Diffusion model alignment using direct preference optimization." In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024.
[0069] To analyze the gradient direction in the reward distribution described above, we define a loss function based on ESM (Explicit Score Matching) (see the following references): • Pascal Vincent. "A connection between score matching and denoising autoencoders." Neural computation, 23(7):1661-1674, 2011. • Chin-Wei Huang et al. "A variational perspective on diffusion-based generative models and score matching." In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021.
[0070]
[0071] Note that the loss function is x unless otherwise specified. t ~q(x t Let's assume |x0) p * (x t The score function of |c) can be decomposed as shown in equation (22).
[0072]
[0073] When converted to a representation in a diffusion model, the reward-based loss function is given by equation (23).
[0074]
[0075] Here, sg[•] is the stop gradient operation, and β is a hyperparameter whose specific selection becomes irrelevant due to gradient normalization. For example, intuitively, equation (23) is the reward gradient ∇ of the model's output direction. xt R(x t This directs the system toward a specific reward design. Below, we will concretize this framework with a concrete reward design and demonstrate that gradients act in tangent space.
[0076] First, to concretize the reward-based framework, we design a specific reward function for element attribution. Based on equation (23), the general reward R(x t Replace the ) with two element-specific rewards. One is the training sample x0 i One reward increases the probability of generating element c, and the other increases the probability of element c. These represent training loss and utility loss, respectively.
[0077] Training sample x0 i To identify the influence of on model generation, we define equation (24).
[0078]
[0079] The term shown in equation (25) within equation (24) is the posterior mean predicted by the diffusion model. This reward is calculated based on the model's training sample x0 i Encourage the generation of samples that are likely to originate from [the specified source].
[0080]
[0081] Following the Diffusion Posterior Sampling (DPS) approach disclosed in the following document, equation (27) is obtained by assuming a Gaussian distribution for the training data, as shown in equation (26). • Hyungjin Chung et al. "Diffusion posterior sampling for general noisy inverse problems." arXiv preprint arXiv:2209.14687, 2022.
[0082]
[0083]
[0084] Thus, in this method, the training loss described above is constructed and used for attribution analysis. The gradient shown in equation (28) acts as a tangent vector on the data manifold (see Chung's literature above) and is consistent with the geometric framework of this method.
[0085]
[0086] Substituting this into equation (23), we obtain the training loss shown in equation (29) below.
[0087]
[0088] Here, equation (30) represents hyperparameters, and their specific selection becomes irrelevant due to gradient normalization.
[0089]
[0090] While DSM loss has a similar objective in that it captures the influence of training samples on generation, the training loss in this method differs in how it constructs the learned signal. For example, DSM provides a reconstruction-driven signal, whereas the DPS-based reward in this method explicitly generates a tangent-space guidance vector. This allows this method to achieve more stable element-level attribution.
[0091] Furthermore, in this method, equation (31) is defined to measure the presence of an element.
[0092]
[0093] Maximizing the reward shown in equation (31) corresponds to maximizing the lower bound of the target element probability (see equation (32) below). Here, p sample This determines the attribution scope. In the case of global attribution, it represents the generation distribution of the model, and in the case of local attribution, it is as shown in equation (33).
[0094]
[0095]
[0096] c targetWhen is a conditioned input, the reward gradient is reduced to a classifier-free guidance vector as shown in equation (34) (see Ho and Salimans below), which functions as an element-relevant tangent vector (see Chung and Brack below). • Jonathan Ho and Tim Salimans. "Classifier-free diffusion guidance." arXiv preprint arXiv:2207.12598, 2022. • Hyungjin Chung et al. "Cfg++: Manifoldconstrained classifier-free guidance for diffusion models." arXiv preprint arXiv:2406.08070, 2024. • Manuel Brack et al. "Sega: Instructing text-to-image models using semantic guidance." Advances in Neural Information Processing Systems, 36:25365-25389, 2023.
[0097]
[0098] For elements embedded in condition c, the contribution of the target element (concept, etc.) within the context is measured using concept slider guidance as shown in equation (35) (see reference below). • Rohit Gandikota et al. "Concept sliders: Lora adaptors for precise control in diffusion models." In European Conference on Computer Vision, pages 172-188. Springer, 2024.
[0099]
[0100] Here, c_ is, for example, a condition to remove from the target element. For example, if c is "pencil drawing of a cat" and c_ is "pencil drawing", then the target element ctarget This becomes "cat". Substituting the above points into the framework of this method (equation (23)), the corresponding utility loss is expressed as in equation (36).
[0101]
[0102] Here... c is a scaling constant, and its value does not affect the final assigned score due to gradient normalization. As mentioned above, for the evaluation function for elements, both the training loss and utility loss are designed to act through the reward gradient in tangent space, so the influence function framework can be applied as shown in equation (37).
[0103]
[0104] This is training sample x0 i The guidance direction induced by and the target element c target This measures consistency with the guidance direction that represents the model. For example, high consistency in the element evaluation function, as shown in equation (37), indicates that the training samples contributed significantly to the model's element generation capability. Thus, the element evaluation function is a function that includes terms corresponding to the elements and terms corresponding to the data being processed. For example, the left-hand side I in equation (37) indicates the degree of influence (contribution) of the data to the target element. That is, the larger the value of I on the left-hand side of equation (37), the higher the influence (contribution) of the data to the target element.
[0105] In Figure 1, the information processing device 100 uses equation (37) to calculate the influence of each training data TD included in the training dataset DS1 when targeting a specific element. The information processing device 100 uses equation (37) to calculate the influence (contribution) of each training data TD. For example, the information processing device 100 estimates that the greater the influence (contribution) calculated using equation (37), the higher the influence (contribution) of that data with respect to the element it targets. For example, the information processing device 100 ranks each training data TD in order of its influence (contribution) to the element, from highest to lowest, by ranking them in descending order of influence (contribution) calculated using equation (37).
[0106] For example, in the case of content CT1, which is an image of a cat in the style of graffiti art, the information processing device 100 sets c to "graffiti art style cat" and sets c_ to "graffiti art style", thereby setting the target element c target The element is designated as "cat". Then, the information processing device 100 calculates the influence of each training data TD included in the training dataset DS1 when the element "cat" is targeted, using equation (37). In this way, the information processing device 100 evaluates the machine learning training data by targeting a specific element. As a result, the information processing system 1 can appropriately evaluate the machine learning training data according to the specification.
[0107] As described above, training data is the data used to train the machine learning model that generated the content. This allows the information processing system 1 to appropriately evaluate the data used to train the machine learning model that generated the content, according to the specified criteria. For example, training data is included in the training dataset stored in memory. This allows the information processing system 1 to appropriately evaluate the data stored in memory, according to the specified criteria.
[0108] Furthermore, for example, an element includes at least one of the concepts and entities contained in the content. This allows the information processing system 1 to appropriately evaluate either the concepts and objects contained in the content according to the specified criteria. In this way, the information processing system 1 can appropriately evaluate various elements contained in the content according to the specified criteria.
[0109] If the content is an image, the element includes at least one element relating to the entire image and at least one element relating to a part of the image. This allows the information processing system 1 to appropriately evaluate any element relating to the entire image or at least one element relating to a part of the image, according to the specifications. For example, the element includes at least one element relating to the style of the image, at least one element relating to the scene contained in the image, at least one element relating to the objects contained in the image, and at least one element relating to the properties of the objects contained in the image. This allows the information processing system 1 to appropriately evaluate any element relating to the style of the image, at least one element relating to the scene contained in the image, at least one element relating to the objects contained in the image, and at least one element relating to the properties of the objects contained in the image, according to the specifications.
[0110] Note that the above process is merely an example, and the information processing system 1 may perform various other processes besides those described above. For example, in order to perform a more stable attribution analysis, the information processing system 1 may apply a deterministic DDIM (Denoising Diffusion Implicit Model) inverse analysis to remove the stochasticity from the forward diffusion process shown in equation (38), and then train sample x0 i From this, we derive a deterministic noisy latent variable as shown in equation (39).
[0111]
[0112]
[0113] The method based on equation (39) described above, when combined with a loss function, can eliminate all randomness from gradient calculations and achieve more stable influence estimation through improved gradient fidelity.
[0114] Furthermore, the sample distribution p in the utility loss (equation (36)) described above. sample The attribution scope can be determined by the selection. For example, the distinction between global attribution and local attribution can be implemented as follows.
[0115] For example, global attribution uses a fully conditional distribution and DDIM from the noise to x t Sample the sample. Local attribution is calculated using the sample distribution p. sample We constrain x0 as shown in equation (40) and require the condition shown in equation (41). Such constraints are given by x0 test x is generated from the noise used for generation by DDIM t This is enforced by sampling.
[0116]
[0117]
[0118] Furthermore, if the magnitude of the loss differs at each time step, a particular gradient may dominate the attribution result. To address this problem, the gradient g of each time step t This is normalized to the unit norm as shown in equation (42).
[0119]
[0120] This allows the information processing system 1 to suppress the excessive influence of a single time step on the final attribution score. Through this normalization, the information processing system 1 can reduce β and σ. data This ensures invariance with respect to hyperparameters, further improving the robustness of the method.
[0121] Furthermore, the information processing system 1 may project the gradient to a lower-dimensional space (k≪d) in accordance with the TRAK described above in order to improve computational efficiency. In addition, by using pre-calculated learned gradients, the information processing system 1 can approximate the Hessian matrix using the Fisher information matrix, thereby suppressing an increase in overhead.
[0122] The information processing system 1 can appropriately evaluate machine learning training data according to specifications using Concept-TRAK, which integrates a loss function based on reward optimization, deterministic sampling, and gradient normalization using an influence function (equation (37)).
[0123] <1-3. Configuration of the Information Processing System According to the Embodiment> Next, the configuration of the information processing system 1, which is an example of an information processing system that performs the information processing according to the embodiment, will be described. Figure 3 is a diagram showing an example of the configuration of the information processing system according to the embodiment of this disclosure. For example, the information processing system 1 shown in Figure 3 is an example of an information processing system that performs the information processing described above.
[0124] As shown in Figure 3, the information processing system 1 includes at least an information processing device 100. The information processing device 100 is connected to other computers via a predetermined communication network N, either by wired or wireless means. Note that the configuration of the information processing system 1 shown in Figure 3 is just one example; any device configuration can be adopted as long as the information processing system 1 includes at least an information processing device 100.
[0125] For example, the information processing system 1 may include multiple information processing devices 100. The information processing system 1 may also include terminal devices used by users who request processing from the information processing devices 100. For example, if the information processing device 100 is a server device, the information processing system 1 includes terminal devices used by users who receive evaluation services provided by the information processing device 100. In this case, the information processing device 100 receives information from the terminal device used by the user requesting evaluation of training data, performs evaluation processing of the training data based on the received information, and transmits the results of the performed evaluation processing of the training data to the terminal device used by the user.
[0126] As shown in Figure 3, the information processing device 100 includes a communication unit 11, an input unit 12, an output unit 13, a storage unit 14, and a control unit 15. In the example in Figure 3, the information processing device 100 includes an input unit 12 (for example, a keyboard or mouse) that receives various operations from the administrator of the information processing device 100, and an output unit 13 (for example, a liquid crystal display) that outputs (displays, etc.) various information.
[0127] The communication unit 11 is implemented, for example, by a NIC (Network Interface Card) or a communication circuit. The communication unit 11 is connected to a communication network N (such as the Internet) by wire or wireless connection and transmits and receives information with other devices via the communication network N.
[0128] The input unit 12 receives various operations from the user. The input unit 12 accepts input from the user. The input unit 12 accepts input from the user of information used to generate a mesh model. The input unit 12 may also accept various operations from the user via a keyboard, mouse, or touch panel provided on the information processing device 100.
[0129] The output unit 13 outputs various types of information. The output unit 13 has a display device (display unit) such as a display and displays various types of information. The output unit 13 displays the information generated by the generation unit 152. The output unit 13 displays the evaluation results from the evaluation unit 153.
[0130] Furthermore, the output unit 13 may have functions to output information in various ways, not limited to display functions. For example, the output unit 13 may have a function to output information as sound. For example, the output unit 13 may have an audio output unit such as a speaker that outputs sound.
[0131] The memory unit 14 is implemented by, for example, semiconductor memory elements such as RAM (Random Access Memory) or flash memory, or by a storage medium (storage device) such as a hard disk or optical disc. The memory unit 14 has a model information storage unit 141 and a training data storage unit 142.
[0132] The model information storage unit 141 according to this embodiment stores information about a model. For example, the model information storage unit 141 stores information (model data) that shows the structure of a model (network). The model information storage unit 141 includes items such as "model ID," "purpose," and "model data."
[0133] The "Model ID" indicates identification information for identifying the model. The "Purpose" indicates the purpose of the corresponding model. The "Model Data" indicates the data of the model. For example, the "Model Data" includes various pieces of information that make up the model, such as information about the network included in the model and functions.
[0134] The model ID "M1" that identifies the content generation model M1 is associated with and stored information indicating that its purpose is "content generation." This indicates that the content generation model M1 is a model used for the task of content generation. In addition, the model ID "M1" that identifies the content generation model M1 is associated with and stored information indicating that the model data of the content generation model M1 is model data MDT1.
[0135] The model information storage unit 141 is not limited to the above and may store various types of information depending on the purpose. For example, the model information storage unit 141 stores parameter information of a model learned (generated) through a learning process.
[0136] The training data storage unit 142 according to this embodiment stores information about the training data. For example, the training data storage unit 142 stores a training dataset containing multiple training data used to train the model.
[0137] The training data storage unit 142 stores the data (training dataset) used for training each model, associating it with the model to be trained. For example, the training data storage unit 142 includes items such as "Target Model ID," "Dataset ID," "Data ID," and "Data."
[0138] The "Target Model ID" indicates the identification information used to identify the model to be trained (the target model). The "Dataset ID" indicates the identification information used to identify the dataset. The "Data ID" indicates the identification information used to identify the data used in the training process of the target model. Furthermore, "Data" refers to the data identified by the Data ID.
[0139] For example, the target model ID "M1" that identifies the content generation model M1 is associated with the dataset ID "DS1" that identifies the dataset DS1 used for its training and is stored accordingly. In addition, the dataset ID "DS1" that identifies the dataset DS1 is associated with the data IDs that identify each training data TD1 to TD6, etc., included in that dataset DS1 and is stored accordingly.
[0140] Furthermore, the training data storage unit 142 may store various types of information depending on the purpose, not limited to those described above.
[0141] Furthermore, the memory unit 14 stores various other types of information. For example, the memory unit 14 stores various types of information used for processing. The memory unit 14 stores specification information that specifies elements included in the content. For example, the memory unit 14 stores specification information that indicates elements specified by the user. For example, the memory unit 14 stores information generated by the generation unit 152. For example, the memory unit 14 stores the generated content. The memory unit 14 stores information that indicates the evaluation results of the training data generated by the evaluation unit 153.
[0142] The control unit 15 has a processing circuit. The control unit 15 is realized by executing a program (for example, the information processing program according to this disclosure) stored inside the information processing device 100 using RAM or the like as a working area, for example, by a CPU (Central Processing Unit) or MPU (Micro Processing Unit). The control unit 15 may also be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).
[0143] As shown in Figure 3, the control unit 15 includes an acquisition unit 151, a generation unit 152, an evaluation unit 153, and a provision unit 154, and realizes or executes the information processing functions and operations described below. Note that the internal configuration of the control unit 15 is not limited to the configuration shown in Figure 3, and other configurations are also acceptable as long as they perform the information processing described later.
[0144] The acquisition unit 151 acquires various types of information. The acquisition unit 151 acquires various types of information from an external information processing device. The acquisition unit 151 acquires various types of information from the storage unit 14. The acquisition unit 151 acquires information received by the input unit 12. For example, the acquisition unit 151 acquires various specifications made by the user received via the input unit 12. The acquisition unit 151 acquires specification information received by the input unit 12. The acquisition unit 151 acquires prompts received by the input unit 12.
[0145] The acquisition unit 151 acquires various information from the storage unit 14. The acquisition unit 151 acquires the content generation model M1 from the model information storage unit 141. The acquisition unit 151 acquires training data from the training data storage unit 142. The acquisition unit 151 receives information via the communication unit 11. The acquisition unit 151 acquires specification information to specify elements.
[0146] The acquisition unit 151 acquires content generated by the machine learning model. The acquisition unit 151 acquires specification information that specifies the elements included in the content. The acquisition unit 151 acquires specification information that specifies an element that includes at least one of the concepts and entities included in the content.
[0147] The acquisition unit 151 acquires the generated image, which is image data generated by the machine learning model, as content. The acquisition unit 151 also acquires specification information that specifies the elements included in the generated image.
[0148] The acquisition unit 151 acquires designation information that specifies an element which includes at least one of the elements relating to the entire raw image or an element relating to a part of the image. The acquisition unit 151 also acquires designation information that specifies an element which includes at least one of the elements relating to the style of the image, an element relating to the scene contained in the image, an element relating to the object contained in the image, and an element relating to the properties of the object contained in the image.
[0149] The generation unit 152 performs processing related to the generation of information. The generation unit 152 performs generation processing using the information stored in the storage unit 14. The generation unit 152 performs generation processing using the information acquired by the acquisition unit 151.
[0150] The generation unit 152 generates content using the content generation model M1. The generation unit 152 generates content by inputting a prompt to the content generation model M1 to instruct it to generate content, and causing the content generation model M1 to output content.
[0151] The evaluation unit 153 performs evaluation processing to evaluate various types of information. The evaluation unit 153 performs evaluation processing using the information generated by the generation processing of the generation unit 152. The evaluation unit 153 performs evaluation processing using the information stored in the storage unit 14.
[0152] The evaluation unit 153 evaluates the machine learning training data targeting the elements specified by the specified information. The evaluation unit 153 evaluates the training data, which is the data used to train the machine learning model that generated the content. The evaluation unit 153 evaluates the training data included in the training dataset stored in the memory unit 120.
[0153] The evaluation unit 153 evaluates the training data using an influence function that targets elements included in the content, based on a method for training data attribution. The evaluation unit 153 evaluates the training data using an influence function that includes terms corresponding to elements and terms corresponding to the data to be processed.
[0154] The evaluation unit 153 may generate various types of information related to the evaluation. The evaluation unit 153 may generate information indicating the evaluation results to be displayed. For example, the evaluation unit 153 may generate information indicating the evaluation results to be provided to the user. For example, the evaluation unit 153 may generate a screen (display content) showing the evaluation results to be displayed on a display device or the like. In this case, the evaluation unit 153 generates the display content using various conventional technologies related to screen display as appropriate. The evaluation unit 153 generates the display content using various conventional technologies related to GUI (Graphical User Interface) as appropriate. For example, the evaluation unit 153 may generate the display content using CSS (Cascading Style Sheets), JavaScript (registered trademark), HTML (HyperText Markup Language), or any language capable of describing the information processing such as the information display and operation reception described above.
[0155] The provisioning unit 154 provides various types of information. The provisioning unit 154 provides various types of information to an external information processing device. The provisioning unit 154 transmits various types of information. The provisioning unit 154 transmits various types of information to an external information processing device. The provisioning unit 154 transmits information stored in the storage unit 14. The provisioning unit 154 transmits information stored in the model information storage unit 141.
[0156] The providing unit 154 transmits the information generated by the generation unit 152. The providing unit 154 transmits the content generated by the generation unit 152 to an external device. For example, the providing unit 154 transmits the content generated by the generation unit 152 to a terminal device used by a user. The providing unit 154 transmits the evaluation results from the evaluation unit 153 to an external device. The providing unit 154 transmits information indicating the evaluation results generated by the evaluation unit 153 to an external device. For example, the providing unit 154 transmits display content indicating the evaluation results generated by the evaluation unit 153 to a terminal device used by a user.
[0157] <1-4. Information Processing Procedure According to the Embodiment> The information processing procedure according to the embodiment will be described below. First, the information processing procedure according to the embodiment will be described using Figure 4. Figure 4 is a flowchart of the information processing procedure according to the embodiment of this disclosure.
[0158] As shown in Figure 4, the information processing system 1 receives the content generated by the machine learning model (step S101). For example, the information processing device 100 acquires the content generated by the machine learning model.
[0159] Then, the information processing system 1 receives specification information that specifies the elements included in the content (step S102). For example, the information processing device 100 acquires specification information that specifies the elements included in the content.
[0160] Then, the information processing system 1 evaluates the machine learning training data targeting the elements specified by the specified information (step S103). For example, the information processing device 100 evaluates the machine learning training data targeting the elements specified by the specified information.
[0161] <1-5. Other Processing Examples (Music, etc.)> The processing described above is merely one example, and the information processing system 1 may perform various other processes besides those described above. Several examples of this are given below.
[0162] For example, in the application examples described above, images were used as an example of content, but the processing described above can be applied to various types of content, not just images. For example, the processing described above may be applied to music data (also simply called "music"), as shown in Figure 5. Figure 5 is a diagram illustrating another example of application. For example, Figure 5 is a diagram illustrating an example where the content is music. Note that the example shown in Figure 5 is the same as the image case described above, except that the content is music, so a detailed explanation of information processing, etc., will be omitted.
[0163] Figure 5 shows an overview of the learning and inference process for a content generation model when the content is music. For example, Figure 5 shows the elements included in the training dataset used to train the content generation model M2 and an image of the content generated by the content generation model M2.
[0164] As shown in Figure 5, the content generation model M2 is trained using a training dataset DS2 that contains multiple training data, each of which is music (e.g., song data) (corresponding to "Training" in Figure 5). For example, each training data in the training dataset DS2 may include information indicating the compositional and arrangement tendencies of a particular musician, such as a (human) voice, an instrument (sound), or a voice. For example, each training data in the training dataset DS2 may include information indicating at least one element from among voice, instrument, composition, arrangement, etc.
[0165] Thus, each training data used to train the content generation model M2 includes at least one element related to musical style, an element related to vocals in the music, and an element related to instruments in the music. For example, each training data used to train the content generation model M2 includes an element related to musical style, which includes at least one of the following: composition of music or arrangement method of music. Figure 5 illustrates an example of the elements included in each training data in the training dataset DS2, but the training dataset DS2 contains a large number of music data (e.g., millions, hundreds of millions, etc.) as training data.
[0166] The learning of the content generation model M2 may be performed by the information processing device 100, as with the content generation model M1, or by a device other than the information processing device 100 (such as a learning device). When a device other than the information processing device 100 (such as a learning device) learns the content generation model M2, the information processing device 100 acquires the content generation model M2 by receiving it from the device that learned the content generation model M2.
[0167] As shown in Figure 5, the content generation model M2 is used to generate content such as content CT2 (corresponding to "Inferece" in Figure 5). For example, in Figure 1, the content generation model M2 outputs content CT2, which is music data, when it receives a prompt instructing it to generate music data.
[0168] The information processing device 100 then performs a process to evaluate each training data included in the training dataset DS2 used to train the content generation model M2, with respect to the elements included in content CT2. For example, the information processing device 100 performs a process to evaluate each training data included in the training dataset DS2 with respect to the elements related to musical instruments included in content CT2. For example, the information processing device 100 performs a process to evaluate each training data included in the training dataset DS2 with respect to the guitar, which is an element related to musical instruments included in content CT2.
[0169] For example, the information processing device 100 uses an influence function (element evaluation function) to evaluate the training data used to train the machine learning model according to the elements, and performs a process to evaluate each training data included in the training dataset DS2 with respect to the element "guitar". For example, the information processing device 100 uses an element evaluation function as shown in equation (37) to evaluate each training data included in the training dataset DS2 with respect to the element "guitar".
[0170] As described above, if the content is music, the information processing device 100 acquires the generated music, which is music data generated by the machine learning model, as the content. The information processing device 100 acquires specification information that specifies the elements included in the generated music. The information processing device 100 acquires specification information that specifies an element which includes at least one of the elements relating to the entire music or the elements relating to a part of the music.
[0171] The information processing device 100 acquires designation information that specifies an element which includes at least one of the following: an element relating to musical style, an element relating to vocals included in the music, and an element relating to musical instruments included in the music. The information processing device 100 also acquires designation information that specifies an element which includes at least one of the following: a composition for the music, or a method of arranging the music, for the element relating to musical style.
[0172] The images and music mentioned above are merely examples of content, and the processing described above may be applied to various other types of content, such as video, text, and ebooks.
[0173] As described above, when the content is music, the elements include at least one element relating to the entire piece of music and at least one element relating to a part of the piece of music. This allows the information processing system 1 to appropriately evaluate either the elements relating to the entire piece of music or the elements relating to a part of the music, according to the specifications.
[0174] For example, an element includes at least one of the following: an element relating to musical style, an element relating to vocals included in the music, and an element relating to instruments included in the music. This allows the information processing system 1 to appropriately evaluate any of the elements relating to musical style, vocals included in the music, or instruments included in the music, according to the specified criteria.
[0175] Furthermore, for example, elements relating to musical style include at least one of the following: musical composition or musical arrangement methods. This allows the information processing system 1 to appropriately evaluate either musical composition or musical arrangement methods according to the specified criteria.
[0176] Furthermore, the information processing system 1 may use a model to evaluate the content (training data). For example, it may use a second model (also called an "element evaluation model") that is different from the first model that generates the content (content generation models M1, M2, etc.) to evaluate the training data. For example, the information processing system 1 may use an LLM (Large Language Model) as the second model (element evaluation model) to evaluate the training data.
[0177] When evaluating training data using a second model, the element evaluation model, for example, the information processing device 100 inputs information into the element evaluation model, causing the element evaluation model to output the evaluation results of the training data. In this way, the information processing device 100 evaluates the training data using the element evaluation model. For example, the information processing device 100 evaluates the training data by inputting a prompt to the element evaluation model instructing it to evaluate the training data, and causing the element evaluation model to output information corresponding to the prompt.
[0178] The following describes a processing example using the content generation model M1 as an example. For example, the information processing device 100 generates information indicating the evaluation of the training data by inputting the content (image) generated by the content generation model M1, the training dataset DS1 used to train the content generation model M1, and a prompt containing specification information for specifying elements into the element evaluation model.
[0179] For example, the information processing device 100 generates information indicating the evaluation of the training data by inputting a prompt to the element evaluation model that includes instruction information such as, "Based on the input information, evaluate each data in the training dataset for the specified element." The instruction information may also include instructions for generating information that ranks each data, such as, "Generate a ranking of each data in the training dataset for the specified element, ordered from most influential to least influential."
[0180] <2. Others> The above-described processes are merely examples, and the information processing device 100 may perform various processes. For example, the information processing device 100 may perform a learning process to learn various models such as content generation models M1 and M2. In this case, the control unit 15 of the information processing device 100 may include a learning unit that performs the learning process. Also, for example, the storage unit 14 of the information processing device 100 may store data used for the learning process (learning data).
[0181] The learning unit of the information processing device 100 (specifically the control unit 15) executes a learning process to learn various models. The learning unit learns various information based on information from an external information processing device and information stored in the storage unit 14. The learning unit stores the model generated by the learning process in the model information storage unit 141. The learning unit also stores the model with updated parameters from the learning process in the model information storage unit 141.
[0182] The learning unit learns various types of information based on the information acquired by the acquisition unit 151. The learning unit learns (generates) a model. The learning unit learns various types of information, such as the model. The learning unit generates a model through learning. The learning unit learns the model using various machine learning techniques. For example, the learning unit learns the parameters of the model (network). The learning unit learns the model using various machine learning techniques.
[0183] The learning unit generates a content generation model M1. The learning unit learns the network parameters. For example, the learning unit learns the network parameters of the content generation model M1.
[0184] The learning unit performs learning processing based on the learning data (training data) stored in the memory unit 14. The learning unit generates a content generation model M1 by performing learning processing using the learning data stored in the memory unit 14. For example, the learning unit generates a model used for content generation processing. The learning unit generates a content generation model M1 by learning the network parameters of the content generation model M1. Note that the learning method (learning processing) is not limited to the method described above, and any known technology can be applied.
[0185] Furthermore, the processing described in the above-described embodiments may be carried out in various other forms (modifications) besides those described above. For example, the system configuration may be various forms, not limited to the examples described above.
[0186] For example, an information processing system may include a device that performs learning processing (learning device) and a device that performs inference processing using the model learned by the learning device (e.g., an information processing device 100). In this case, the information processing system may include both the learning device and the information processing device 100. The above is just one example, and the information processing system may be implemented in various configurations. For example, the information processing device 100 may be a device that performs both learning processing and inference processing. That is, in an information processing system, the learning device and the inference device may be integrated.
[0187] Furthermore, among the processes described in each of the above embodiments, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, specific names, and information including various data and parameters shown in the above document and drawings can be changed at will unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown.
[0188] Furthermore, each element of the illustrated device is a functional concept and does not necessarily have to be physically configured as shown. In other words, the specific forms of distribution and integration of each device are not limited to those shown, and all or part of them can be functionally or physically distributed and integrated in any unit according to various loads and usage conditions.
[0189] Furthermore, the embodiments and modifications described above can be combined as appropriate, provided that the processing content is not inconsistent.
[0190] Furthermore, the effects described herein are merely illustrative and not limiting; other effects may also occur.
[0191] <3. Hardware Configuration> The information processing device 100 and the like according to the embodiments described above are realized by a computer 1000 having a configuration such as that shown in Figure 6. The information processing device 100 will be used as an example for explanation. Figure 6 is a hardware configuration diagram showing an example of a computer that realizes the functions of the information processing device. The computer 1000 has a processing circuitry 1100, RAM 1200, ROM 1300, secondary storage device 1400, communication interface 1500, input / output interface 1600, display unit 1700, camera unit 1800, microphone 1900, and speaker 2000. The various parts of the computer 1000 are connected by a bus 1050.
[0192] The processing circuit 1100 operates based on a program stored in the ROM 1300 or secondary storage device 1400, and controls each part. For example, the processing circuit 1100 loads the program stored in the ROM 1300 or secondary storage device 1400 into the RAM 1200 and executes processing corresponding to various programs.
[0193] ROM 1300 stores boot programs such as the BIOS (Basic Input Output System) that are executed by the processing circuit 1100 when the computer 1000 starts up, as well as programs that depend on the computer 1000's hardware.
[0194] The secondary storage device 1400 is a computer-readable recording medium that non-temporarily stores programs executed by the processing circuit 1100 and data used by such programs. Specifically, the secondary storage device 1400 is a recording medium that stores programs for each process of the information processing device 100 according to the embodiment, which is an example of program data 1450.
[0195] The communication interface 1500 is an interface for the computer 1000 to connect to the external network 1550. The communication interface 1500 corresponds to the communication unit 11 of the information processing device 100. For example, the processing circuit 1100 receives data from other devices or transmits data generated by the processing circuit 1100 to other devices via the communication interface 1500.
[0196] The input / output interface 1600 is an interface for connecting the input / output device 1650 and the computer 1000. For example, the processing circuit 1100 receives data from input devices such as a microphone 1900 or a touch panel via the input / output interface 1600. The processing circuit 1100 also transmits data to output devices such as a display unit 1700 or a speaker 2000 via the input / output interface 1600. The input / output interface 1600 may also function as a media interface for reading programs recorded on a predetermined recording medium (media). Examples of media include optical recording media such as DVDs (Digital Versatile Discs) and PDs (Phase Change Rewritable Disks), magneto-optical recording media such as MOs (Magneto-Optical Disks), tape media, magnetic recording media, or semiconductor memory.
[0197] The display unit 1700 is an interface for displaying information processed by the computer 1000. The display unit 1700 is, for example, a liquid crystal display or an organic electroluminescent display (Organic Electro Luminescence Display). Alternatively, the display unit 1700 may be a touch panel display device or an image projection device.
[0198] The camera unit 1800 is an interface for the computer 1000 to capture images. The microphone 1900 is an interface for the computer 1000 to capture sound. The speaker 2000 is an interface for the computer 1000 to output processed sound. The various parts of the computer 1000 are connected by the bus 1050. Each interface does not necessarily have to be located inside the computer 1000, but may be located outside the computer 1000 via a network or the like. Furthermore, each part of the computer 1000 may be controlled by a circuit different from the processing circuit 1100. For example, the display unit 1700 may be controlled not by the processing circuit 1100, but by a circuit dedicated to display processing provided within the display unit 1700.
[0199] For example, when the computer 1000 functions as an information processing device 100 according to the embodiment, the processing circuit 1100 of the computer 1000 functions as a control unit 15 by executing a program loaded onto the RAM 1200. The secondary storage device 1400 stores the information processing program according to this disclosure and various data stored by the storage unit 14. The processing circuit 1100 reads and executes the program data 1450 from the secondary storage device 1400, but as another example, these programs may be obtained from other devices via an external network 1550. In other words, the secondary storage device 1400 is not limited to being inside the computer 1000, but may be located outside the computer 1000. The processing circuit 1100 is an example of an integrated circuit, and CPU, MPU, GPU, APU, ASIC, and FPGA can all be considered integrated circuits.
[0200] Furthermore, this technology can also take the following configurations: (1) An information processing system including a processing circuit, wherein the processing circuit receives content generated by a machine learning model, receives designation information that specifies elements included in the content, and evaluates machine learning training data targeting the elements specified by the designation information. (2) The information processing system according to (1), wherein the training data is data used to train the machine learning model that generated the content. (3) The information processing system according to (2), wherein the training data is included in a training dataset stored in memory. (4) The information processing system according to any one of (1) to (3), wherein the element includes at least one of a concept included in the content and an entity included in the content. (5) The information processing system according to any one of (1) to (4), wherein the generated image, which is image data generated by the machine learning model, is received as the content, and receives designation information that specifies the elements included in the generated image. (6) The information processing system according to (5), wherein the element includes at least one of an element relating to the entire image and an element relating to a part of the image. (7) The information processing system according to (6), wherein the element includes at least one of the following: an element relating to the style of an image, an element relating to a scene contained in an image, an element relating to an object contained in an image, and an element relating to the properties of an object contained in an image. (8) The information processing system according to any one of (1) to (4), wherein the element receives generated music, which is music data generated by the machine learning model, as the content, and receives the designation information that specifies the elements contained in the generated music. (9) The information processing system according to (8), wherein the element includes at least one of the following: an element relating to the whole of the music, an element relating to a part of the music. (10) The information processing system according to (9), wherein the element includes at least one of the following: an element relating to the style of the music, an element relating to a vocal contained in the music, and an element relating to an instrument contained in the music.(11) The information processing system according to (10), wherein the elements relating to the style of music include at least one of composition for music or arrangement methods for music. (12) The information processing system according to any one of (1) to (11), wherein the training data is evaluated using a machine learning model different from the machine learning model. (13) The information processing system according to any one of (1) to (12), wherein the training data is evaluated using an influence function that targets the elements included in the content, based on a method relating to training data attribution. (14) The information processing system according to (13), wherein the influence function is a function that includes a term corresponding to the element and a term corresponding to the data to be processed. (15) An information processing method that receives content generated by a machine learning model, receives designation information that specifies elements included in the content, and evaluates machine learning training data targeting the elements specified by the designation information. (16) A computer-readable storage medium which stores a program that receives content generated by a machine learning model, receives designation information that specifies elements included in the content, and causes a computer to perform a process of evaluating machine learning training data with respect to the elements specified by the designation information.
[0201] 1. Information Processing System 100. Information Processing Device 11. Communication Unit 12. Input Unit 13. Output Unit 14. Storage Unit 141. Model Information Storage Unit 142. Training Data Storage Unit 15. Control Unit 151. Acquisition Unit 152. Generation Unit 153. Evaluation Unit 154. Provision Unit
Claims
1. An information processing system including a processing circuit, wherein the processing circuit receives content generated by a machine learning model, receives designation information that specifies elements included in the content, and evaluates machine learning training data with respect to the elements specified by the designation information.
2. The information processing system according to claim 1, wherein the training data is data used to train the machine learning model that generated the content.
3. The information processing system according to claim 2, wherein the training data is included in a training dataset stored in memory.
4. The information processing system according to claim 1, wherein the element comprises at least one of the concepts and entities included in the content.
5. The information processing system according to claim 1, which receives a generated image, which is image data generated by the machine learning model, as the content, and receives the designation information that specifies the elements included in the generated image.
6. The information processing system according to claim 5, wherein the element comprises at least one of an element relating to the entire image and an element relating to a part of the image.
7. The information processing system according to claim 6, wherein the element comprises at least one of the following: an element relating to the style of an image, an element relating to a scene contained in an image, an element relating to an object contained in an image, and an element relating to the properties of an object contained in an image.
8. The information processing system according to claim 1, which receives generated music, which is music data generated by the machine learning model, as content, and receives the designation information that specifies the elements included in the generated music.
9. The information processing system according to claim 8, wherein the element comprises at least one of the elements relating to the entire piece of music and the elements relating to a part of the piece of music.
10. The information processing system according to claim 9, wherein the elements include at least one of the following: elements relating to musical style, elements relating to vocals included in music, and elements relating to musical instruments included in music.
11. The information processing system according to claim 10, wherein the elements relating to musical style include at least one of composing music or arranging music.
12. The information processing system according to claim 1, which evaluates the training data using a machine learning model different from the machine learning model described above.
13. An information processing system according to claim 1, which evaluates the training data using an influence function that targets the elements included in the content, based on a method for training data attribution.
14. The information processing system according to claim 13, wherein the influence function is a function that includes a term corresponding to the element and a term corresponding to the data to be processed.
15. An information processing method that receives content generated by a machine learning model, receives designation information that specifies elements included in the content, and evaluates machine learning training data targeting the elements specified by the designation information.
16. A computer-readable storage medium containing a program that receives content generated by a machine learning model, receives designation information specifying elements included in the content, and causes a computer to perform a process of evaluating machine learning training data targeting the elements specified by the designation information.
