Information processing system, information processing method, and program

Abstract

To facilitate analysis of a prediction basis without depending on a presentation mode of XAI.SOLUTION: An AI system includes a prediction rationale acquiring unit that acquires a prediction rationale indicating a basis of prediction by an AI model, a prompt generating unit that generates a prompt that causes a generation AI to generate an analysis result corresponding to the prediction rationale from the prediction rationale acquired by the prediction rationale acquiring unit, a report generating unit that generates a report by inputting the prompt generated by the prompt generating unit to the generation AI, and a report outputting unit that outputs the analysis result generated by the report generating unit.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing system, an information processing method, and a program. [Background technology]

[0002] There is a technology called XAI that presents the basis for predictions made by an AI predictive model. Using XAI makes it possible to visualize the basis on which the predictive model made its prediction. It is expected that such prediction basis can be used to plan measures and improve training data and features. However, analyzing the prediction basis is difficult, and it is difficult to say that it is fully understood and utilized in some cases. To address this issue, Patent Document 1 discloses a technology that improves the interpretability of analysis results by XAI for graph-structured data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2023-118076 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the technology in Patent Document 1 is only applicable to graph-structured data and cannot be applied to other data. There are various types of XAI, including SHAP (SHapley Additive exPlanation), GradCam (Gradient-weighted Class Activation Mapping), and Influence Function. SHAP displays the contribution of features to predicted values ​​using bar graphs, etc. GradCam displays image regions that are of interest in predictions such as image recognition. Influence Function calculates the degree to which individual learning data has influenced the prediction results. It would be better if prediction grounds presented in such various ways could be easily analyzed.

[0005] The present invention has been made in consideration of this situation, and aims to provide an information processing system, information processing method, and program that can facilitate analysis of prediction basis regardless of the presentation format of XAI. [Means for solving the problem]

[0006] The information processing system of the present invention includes a prediction basis acquisition unit that acquires prediction basis indicating the basis for a prediction made by an AI model; a prompt generation unit that generates a prompt from the prediction basis acquired by the prediction basis acquisition unit to cause a generation AI to generate an analysis result corresponding to the prediction basis; a report generation unit that generates an analysis result of the prediction basis by inputting the prompt generated by the prompt generation unit into the generation AI; and a report output unit that outputs the analysis result generated by the report generation unit.

[0007] The information processing method of the present invention is an information processing method performed by a computer, in which a prediction basis acquisition unit acquires prediction basis indicating the basis for a prediction made by an AI model, a prompt generation unit generates a prompt from the prediction basis acquired by the prediction basis acquisition unit to cause a generation AI to generate an analysis result corresponding to the prediction basis, a report generation unit generates an analysis result of the prediction basis by inputting the prompt generated by the prompt generation unit to the generation AI, and a report output unit outputs the analysis result generated by the report generation unit.

[0008] The program of the present invention causes a computer to acquire prediction grounds that indicate the grounds for predictions made by an AI model, generate a prompt from the prediction grounds that causes a generation AI to generate an analysis result corresponding to the prediction grounds, input the generated prompt into the generation AI to generate an analysis result of the prediction grounds, and output the analysis result. [Effects of the Invention]

[0009] According to the present invention, it is possible to facilitate analysis of prediction grounds regardless of the presentation format of XAI. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a block diagram showing an example of the configuration of an information processing system 1 according to an embodiment. [Figure 2] FIG. 2 is a sequence diagram illustrating the flow of processing performed by the information processing system 1 according to the embodiment. [Figure 3] FIG. 2 is a sequence diagram illustrating the flow of processing performed by the information processing system 1 according to the embodiment. [Figure 4] FIG. 2 is a sequence diagram illustrating the flow of processing performed by the information processing system 1 according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0012] (About Information Processing System 1) 1 is a block diagram showing an example of the configuration of an information processing system 1 according to an embodiment. The information processing system 1 is a system that analyzes the prediction basis of an AI model.

[0013] Generally, the prediction process using an AI model is a black box, making it difficult for humans to interpret or understand. In recent years, XAI has come to be used to explain the basis for predictions made by AI models. However, the basis for XAI predictions can be difficult to understand, and depending on the target being predicted by the AI ​​model, understanding the basis for the prediction may require business knowledge or specialized knowledge. For this reason, when trying to plan measures to improve the accuracy of an AI model's predictions using the basis for XAI predictions, it is necessary to interpret the basis for the prediction while aligning it with business knowledge and specialized knowledge, which is a specialized, time-consuming, and labor-intensive task. Furthermore, because the interpretation of the basis for predictions varies depending on the person interpreting it, measures can vary in effectiveness, resulting in limited effectiveness or ineffective measures.

[0014] To address this issue, the information processing system 1 of this embodiment uses a generation AI to analyze the prediction basis calculated using the XAI. The generation AI is an AI that learns patterns and relationships based on training data and generates content based on the learned content. An LLM (Large Language Model) can be used as the generation AI. Furthermore, for example, GPT-4 (registered trademark), Llama2, PaLM2, etc. can be used as the generation AI. In this embodiment, the prediction basis calculated using XAI is input to the generation AI, and the result of interpreting the input prediction basis (analysis result) is generated as content. The content generated by the generation AI may be text, an image, or a combination of these.

[0015] As shown in FIG. 1, the information processing system 1 includes, for example, an XAI report generation device 10 and a learning device 20. The learning device 20 generates an AI model by training the learning model on the learning data, and calculates the prediction basis of the generated AI model using XAI. The AI ​​model used here may be any model that can calculate the prediction basis using at least XAI, i.e., explainable AI. The XAI report generation device 10 analyzes the prediction basis calculated using the XAI using the generation AI, and outputs an XAI report as the analysis result. In the information processing system 1, an XAI report is presented to a human (analyst M). Here, the XAI report is generated by the generation AI. Therefore, an XAI report can be generated that analyzes the basis for XAI predictions in a way that is easy for humans to understand. In addition, an XAI report that includes explanations of business knowledge and specialized knowledge can be generated as needed. Analyst M plans measures using the XAI report that has been analyzed in a way that is easy for humans to understand. This makes it possible to plan accurate measures in a short amount of time.

[0016] (About the learning device 20) The learning device 20 is a computer. The learning device 20 is realized by, for example, a cloud, a server device, a PC (personal computer), etc. As shown in FIG. 1 , the learning device 20 includes, for example, an accuracy evaluation unit 21, a prediction basis generation unit 22, an AI model storage unit 23, an AI model learning unit 24, and a learning data storage unit 25.

[0017] The following describes an example of an AI model that uses a model that makes predictions using images. For example, the AI ​​model is a model that inputs captured images of plant leaves and predicts the degree to which a lesion is present on the leaves in the input images. In this case, the AI ​​model learns the correspondence between images and the presence or absence of a lesion. Specifically, the AI ​​model learns the correspondence between images and the presence or absence of a lesion by learning training data that associates captured training images of plant leaves with the presence or absence of a lesion on the leaves in the training images. By learning this correspondence, the AI ​​model becomes able to predict the degree to which a lesion is present on the leaves captured in the input images.

[0018] The training data storage unit 25 is a storage medium that stores training data. The training data here is training data in which training images of plant leaves are associated with whether or not the leaves in the training images have a disease. For example, analyst M collects training images (images of plant leaves) that the AI ​​model wants to learn from. Analyst M determines whether the collected training images are images of leaves with a disease or images of leaves without a disease, and generates training data by labeling the determination results. Analyst M stores the newly generated training data in the training data storage unit 25.

[0019] The AI ​​model learning unit 24 causes the AI ​​model to learn the learning data stored in the learning data storage unit 25. For example, the AI ​​model learning unit 24 may cause a learning model (AI model) that has not yet learned anything to learn new learning data, or may perform additional learning in which an already-learned AI model learns additional learning data.

[0020] The accuracy evaluation unit 21 evaluates the accuracy of predictions made by the AI ​​model. Evaluating the accuracy here means determining whether the accuracy of the predictions is equal to or greater than a threshold. The accuracy evaluation unit 21, for example, inputs a verification image to an AI model that has been trained with training data by the AI ​​model training unit 24. The verification image is an image (an image of a plant leaf) that has not been trained by the AI ​​model, and it is known whether or not the leaf has a lesion.

[0021] The accuracy evaluation unit 21 acquires a prediction result output from the AI ​​model in response to inputting a verification image to the AI ​​model. For example, the accuracy evaluation unit 21 compares the prediction result output from the AI ​​model with a correct answer indicating whether or not a lesion has occurred in the leaf captured in the verification image, and evaluates the accuracy of the prediction by the AI ​​model based on the degree to which the prediction result matches the correct answer. For example, the accuracy evaluation unit 21 uses multiple different verification images to calculate the probability that each prediction result matches the correct answer, and determines whether the calculated probability is equal to or greater than a threshold.

[0022] If the accuracy of the prediction by the AI ​​model is equal to or greater than the threshold, the accuracy evaluation unit 21 stores information for constructing the AI ​​model in the AI ​​model storage unit 23.

[0023] On the other hand, if the accuracy of the prediction by the AI ​​model is less than the threshold, the accuracy evaluation unit 21 outputs information indicating the AI ​​model to the prediction basis generation unit 22.

[0024] The prediction basis generation unit 22 acquires information indicating the AI ​​model from the accuracy evaluation unit 21, and calculates the basis for prediction by the AI ​​model based on the acquired information. The prediction basis generation unit 22 calculates the basis for prediction by the AI ​​model using XAI (Explainable AI), that is, explainable AI.

[0025] Here, the XAI used to identify the basis for prediction can be any conventional method, such as SHAP (Shapley Additive exPlanation), GradCam (Gradient-weighted Class Activation Mapping), and TracIn. SHAP is a method that applies the Shapley value from cooperative game theory to machine learning. By using SHAP-based XAI, it is possible to calculate the contribution of each feature to the estimation by the AI ​​model, and furthermore, to quantitatively show whether each feature has had a positive or negative effect on the AI ​​prediction. GradCam is a method for identifying which parts of an image an AI model focuses on when making predictions. Using GradCam, it is possible to identify the areas of an image that contributed to the AI ​​model's prediction, as well as the data that contributed to the prediction. Using GradCam, for example, it is possible to show the degree to which each pixel or area in an image contributed to the AI ​​model's prediction using a heat map or similar. TracIn is a method for tracking the process by which training data is learned by an AI model. By tracking the training process, it is possible to identify whether the training data has influenced the prediction accuracy in a way that improves it, or whether it has influenced the prediction accuracy in a way that decreases it. Using TracIn, it is possible to show the impact that each piece of training data has had on the accuracy of predictions made by an AI model.

[0026] The AI ​​model storage unit 23 stores information for constructing an AI model, such as internal parameter settings according to the configuration of the AI ​​model.

[0027] The memory unit (including the AI ​​model memory unit 23 and the learning data memory unit 25) included in the learning device 20 is configured by a storage medium such as a hard disk drive (HDD), flash memory, electrically erasable programmable read-only memory (EEPROM), random access read / write memory (RAM), read-only memory (ROM), or a combination of these. The memory unit included in the learning device 20 stores programs for executing various processes in the learning device 20 and temporary data used when performing the various processes.

[0028] In addition, the functional units (including the accuracy evaluation unit 21, the prediction basis generation unit 22, and the AI ​​model learning unit 24) of the learning device 20 are realized by executing a program on a CPU (Central Processing Unit) and / or a GPU (Graphics Processing Unit) that the learning device 20 has as hardware.

[0029] Any model may be used as the AI ​​model, for example, a convolutional neural network (CNN), a support vector machine (SVM), decision tree learning, genetic programming, or a model based on a combination of these. Any learning method may be used for learning, for example, prediction may be performed using a trained model trained by machine learning, deep learning, quantum computing, or a combination of these methods.

[0030] The prediction basis generating unit 22 of the information processing system 1 may be configured to select an XAI that calculates the prediction basis according to the AI ​​model. The prediction basis generating unit 22 may use multiple types of XAI to calculate the prediction basis, or may calculate multiple prediction basis using one type of XAI. For example, the prediction basis generation unit 22 applies SHAP to the AI ​​model to identify the learning data and the feature that contributed most to the prediction by the AI ​​model. Furthermore, the prediction basis generation unit 22 may apply GradCam to the AI ​​model (an AI model that makes a prediction using an image) to identify an area in an image input to the AI ​​model that contributed to the prediction by the AI ​​model.

[0031] (About the XAI report generator 10) The XAI report generation device 10 is a computer. The XAI report generation device 10 is realized, for example, by a cloud, a server device, a PC (personal computer), etc. As shown in FIG. 1, the XAI report generation device 10 includes, for example, a prediction basis acquisition unit 11, a prompt generation unit 12, a report generation unit 13, and a report display unit 14.

[0032] The prediction basis acquisition unit 11 acquires prediction basis indicating the basis for prediction by the AI ​​model. For example, the prediction basis acquisition unit 11 acquires prediction basis calculated by the prediction basis generation unit 22 of the learning device 20. The prediction basis acquisition unit 11 outputs the acquired prediction basis to the prompt generation unit 12.

[0033] The prompt generation unit 12 generates a prompt that causes the generation AI to generate an analysis result corresponding to the prediction basis from the prediction basis acquired by the prediction basis acquisition unit 11. The prompt here is information that instructs the generation AI on the content that the user wants to generate. The prompt generating unit 12 generates, as a prompt, a text sentence indicating an instruction such as "Please analyze the prediction basis" together with the prediction basis acquired from the prediction basis acquiring unit 11, for example. Here, the prompt generation unit 12 may include more specific instructions in the " " in the prompt. For example, the prompt may specify the knowledge to be used for analysis, such as "Please analyze the prediction basis based on the business knowledge (expert knowledge) of the target predicted by the AI ​​model." Furthermore, the field of business knowledge may be specified more specifically, such as marketing, manufacturing, medical, education, or finance. In this case, information indicating the field of business knowledge may be set in advance or may be specified by the analyst M. For example, the prompt generation unit 12 uses the information specified by the analyst M to generate a prompt including information specifying the field of business knowledge. Furthermore, the prompt generation unit 12 may specify the purpose for which the analysis results are to be used in the contents of " " in the prompt. For example, the purpose of the analysis may be specified, such as "Please propose measures that can improve the prediction accuracy of the AI ​​model based on the prediction grounds, along with the reasons for doing so." In this case, information indicating the purpose of the analysis may be set in advance or may be specified by the analyst M. For example, the prompt generation unit 12 uses information specified by the analyst M to generate a prompt including information specifying the purpose of the analysis. Furthermore, the prompt generation unit 12 may specify in the contents of " " in the prompt which person will use the analysis results. For example, the target for which the analysis results will be used may be specified, such as "Please explain the basis for the prediction so that it can be understood by a sales person who has little business knowledge (specialized knowledge) about the target predicted by the AI ​​model." In this case, information indicating the target for which the analysis results will be used may be set in advance or may be specified by the analyst M. For example, the prompt generation unit 12 uses the information specified by the analyst M to generate a prompt including information specifying the target for which the analysis results will be used.

[0034] The report generation unit 13 generates an XAI report. The report generation unit 13 generates an XAI report that indicates the analysis results of the prediction basis by inputting the prompt generated by the prompt generation unit 12 into the generation AI. The report display unit 14 displays the XAI report generated by the report generation unit 13 on a display (not shown) of the XAI report generation device 10.

[0035] The storage unit (not shown) included in the XAI report generation device 10 is configured by a storage medium such as a HDD, flash memory, EEPROM, RAM, or ROM, or a combination of these. The storage unit included in the XAI report generation device 10 stores programs for executing various processes in the XAI report generation device 10 and temporary data used when performing various processes.

[0036] In addition, the functional units (including the prediction basis acquisition unit 11, the prompt generation unit 12, the report generation unit 13, and the report display unit 14) provided in the XAI report generation device 10 are realized by executing a program on the CPU and / or GPU provided as hardware in the XAI report generation device 10.

[0037] The analyst M visually checks the XAI report displayed on the display (not shown) of the XAI report generation device 10, considers measures to improve the prediction accuracy of the AI ​​model, generates learning data that reflects the considered matters, and stores the generated learning data in the learning data storage unit 25. In the information processing system 1, learning of the AI ​​model by the learning device 20, calculation of prediction basis, generation of prompts by the XAI report generation device 10, and generation and display of the XAI report are repeatedly executed until the prediction accuracy of the AI ​​model reaches or exceeds a threshold. When the prediction accuracy of the AI ​​model reaches or exceeds a threshold, the prediction basis generation unit 22 stores information for constructing the AI ​​model in the AI ​​model storage unit 23.

[0038] Here, the flow of processing performed by the information processing system 1 will be described with reference to Fig. 2 to Fig. 4. Fig. 2 to Fig. 4 are sequence diagrams for explaining the flow of processing performed by the information processing system 1 according to the embodiment.

[0039] Figure 2 mainly shows the flow of a series of processes for calculating the prediction basis of an AI model. First, learning is performed using learning data. Analyst M instructs the AI ​​model learning unit 24 to perform learning on the AI ​​model (step S10). The AI ​​model learning unit 24 refers to the learning data storage unit 25 and requests learning data (step S11). In response, the learning data is output from the learning data storage unit 25 to the AI ​​model learning unit 24 (step S12). The AI ​​model learning unit 24 causes the AI ​​model to learn the examination image data for learning acquired from the learning data storage unit 25 (step S13). The AI ​​model learning unit 24 inputs the learning data to the AI ​​model and adjusts the internal parameters of the AI ​​model so that the prediction result output from the AI ​​model approaches the judgment result associated with the learning data.

[0040] Next, the prediction accuracy of the AI ​​model that has learned the learning data is evaluated. The AI ​​model learning unit 24 outputs information on the learned model (AI model) that has learned the learning data to the accuracy evaluation unit 21 (step S14). The accuracy evaluation unit 21 evaluates the prediction accuracy of the AI ​​model learned by the AI ​​model learning unit 24 (step S15). The accuracy evaluation unit 21 evaluates the prediction accuracy of the AI ​​model using, for example, verification data.

[0041] Next, XAI is applied to the AI ​​model to calculate prediction basis. If the accuracy of the prediction by the AI ​​model does not achieve the target accuracy, the accuracy evaluation unit 21 outputs information about the AI ​​model, such as internal data settings for constructing the AI ​​model and learning data learned by the AI ​​model, to the prediction basis generation unit 22 (step S16). The prediction basis generation unit 22 calculates prediction basis by the AI ​​model using the information acquired from the accuracy evaluation unit 21. Specifically, the prediction basis generation unit 22 applies GradCam to the AI ​​model to calculate a region that was focused on at the time of prediction in the image input to the AI ​​model at the time of prediction (step S17). Furthermore, the prediction basis generation unit 22 applies SHAP to the AI ​​model to identify the learning data and / or the feature that contributed most to the prediction by the AI ​​model (step S18). Furthermore, the prediction basis generating unit 22 applies TraceIn to the AI ​​model to calculate the training examination image data with improved accuracy and the training examination image data with reduced accuracy in the prediction (step S19). Here, TraceIn is a type of XAI method using an influence function, and is an XAI method that presents training images with improved prediction accuracy and training images with reduced accuracy.

[0042] Then, the calculated prediction basis is output. The prediction basis generating unit 22 outputs the prediction basis calculated in steps S17 to S19 to the prediction basis acquiring unit 11 of the XAI report generating device 10 (step S20).

[0043] FIG. 3 mainly shows the flow of a series of processes for outputting an XAI report. First, a prompt is generated. As shown in FIG. 2, in step S20, the prediction basis is output from the prediction basis generation unit 22 to the prediction basis acquisition unit 11. The prediction basis acquisition unit 11 outputs the prediction basis acquired from the prediction basis generation unit 22 to the prompt generation unit 12 (step S21). Meanwhile, the prediction basis is displayed on a display (not shown) of the XAI report generation device 10. The analyst M visually checks the prediction basis displayed on the display (not shown) and inputs more specific information to be used for analyzing the prediction basis. Examples of more specific information include information specifying the field of business knowledge (specialized knowledge) used to analyze the prediction basis, information indicating the purpose of analyzing the prediction basis, and information specifying the position and knowledge level of the target (team or person belonging to the team) who will use the analysis results of the prediction basis. The information input by the analyst M is acquired by the prompt generation unit 12 via the control unit of the XAI report generation device 10 (step S22). The prompt generator 12 generates a prompt using the prediction basis acquired from the prediction basis acquirer 11 and more specific information used to analyze the prediction basis acquired from the analyst M (step S23).

[0044] Next, an XAI report is generated. The prompt generator 12 outputs information indicating the generated prompt, such as image information indicating an image or graph related to the prediction basis, and text information indicating analysis instructions, to the report generator 13 (step S24). The report generator 13 inputs the prompt acquired from the prompt generator 12 into the generation AI, thereby generating an XAI report as the analysis result of the XAI (step S25).

[0045] Then, the display of the XAI report is executed. The report generation unit 13 outputs image information for displaying the generated XAI report to the report display unit 14 (step S26). The report display unit 14 displays the XAI report on a display (not shown) of the XAI report generation device 10 using the image information acquired from the report generation unit 13 (step S27).

[0046] Figure 4 mainly shows the process flow for retraining the AI ​​model as an improvement measure based on the XAI report. First, an analyst considers and implements improvement measures. As shown in FIG. 3, the report display unit 14 displays the XAI report on the display (not shown) of the XAI report generation device 10 using image information acquired from the report generation unit 13 (step S27). The analyst M visually checks the displayed XAI report and considers improvement measures (step S28). For example, the analyst M generates training data that reflects the improvement measures and stores the generated training data in the training data storage unit 25 (step S29). The analyst M also implements improvement measures by reviewing the clustering of each training data and correcting the class classification (step S30). The analyst M also implements improvement measures such as removing noise contained in the training data and eliminating mislabeling (step S31). The analyst M stores the training data on which the improvement measures have been implemented in the training data storage unit 25.

[0047] Next, re-learning is performed using the learning data that has been improved as a result of the implementation of the improvement measures. Analyst M instructs AI model learning unit 24 to re-learn the AI ​​model using the learning data for which the improvement measures have been implemented (step S32). AI model learning unit 24 acquires learning data for re-learning from learning data storage unit 25, and re-learns the AI ​​model using the acquired learning data (step S33).

[0048] Next, the accuracy of the re-learned AI model is evaluated. The AI ​​model learning unit 24 outputs information about the AI ​​model (trained model) that has been trained using the training data for re-learning to the accuracy evaluation unit 21 (step S34). The accuracy evaluation unit 21 evaluates the accuracy of predictions made by the AI ​​model trained by the AI ​​model learning unit 24 (step S35). The accuracy evaluation unit 21 evaluates the prediction accuracy of the AI ​​model using, for example, verification data.

[0049] Then, when the accuracy reaches a target value, the AI ​​model is stored. When the accuracy of prediction by the AI ​​model reaches the target accuracy, the accuracy evaluation unit 21 outputs information for constructing an AI model (trained model) to the AI ​​model storage unit 23 (step S36). The AI ​​model storage unit 23 stores the information for constructing the AI ​​model acquired from the accuracy evaluation unit 21 (step S37).

[0050] If the accuracy of the prediction by the AI ​​model does not reach the target accuracy in step S35, the accuracy evaluation unit 21 returns to the process shown in step S16, and calculates the prediction basis again and generates an XAI report.

[0051] As described above, the information processing system 1 of the embodiment includes a prediction basis acquisition unit 11, a prompt generation unit 12, a report generation unit 13, and a report display unit 14 (report output unit). The prediction basis acquisition unit 11 acquires prediction basis indicating the basis for a prediction made by an AI model. The prompt generation unit 12 generates a prompt from the prediction basis acquired by the prediction basis acquisition unit 11, causing the generation AI to generate an analysis result corresponding to the prediction basis. The report generation unit 13 generates an analysis result of the prediction basis (an XAI report) by inputting the prompt generated by the prompt generation unit 12 to the generation AI. The report display unit 14 outputs the analysis result (an XAI report) generated by the report generation unit 13. As a result, the information processing system 1 of the embodiment can cause the generation AI to generate an XAI report as the analysis result of the XAI, making it easy to analyze the prediction basis regardless of the presentation format of the XAI.

[0052] Furthermore, in the information processing system 1 of the embodiment, the prediction basis acquisition unit 11 acquires, as the prediction basis, the contribution of the feature amount used by the AI ​​model for prediction, such as that calculated by applying SHAP to the AI ​​model. This makes it easy to consider how the contribution of the feature amount used by the AI ​​model for prediction should be reflected in improvement measures in the information processing system 1 of the embodiment.

[0053] In the information processing system 1 according to the embodiment, the AI ​​model is a model that performs prediction using an image. The prediction basis acquisition unit 11 acquires, as the prediction basis, information indicating which area of ​​the input image the AI ​​model focused on in the prediction using the input image, such as information calculated by applying GradCam to the AI ​​model. This makes it easy to consider how the area focused on in the prediction by the AI ​​model should be reflected in improvement measures in the information processing system 1 according to the embodiment.

[0054] Furthermore, in the information processing system 1 of the embodiment, the prediction basis acquisition unit 11 acquires, as the prediction basis, information indicating the degree of influence that has affected the prediction accuracy of the learning data learned by the AI ​​model, such as calculated by applying an influence function (e.g., TracIn) to the AI ​​model. This makes it easy to consider how the degree of influence that the learning data has on the prediction accuracy should be reflected in improvement measures in the information processing system 1 of the embodiment.

[0055] In the information processing system 1 according to the embodiment, the prompt generator 12 generates a prompt including knowledge about the target predicted by the AI ​​model, such as industry knowledge or specialized knowledge about the target. This allows the information processing system 1 according to the embodiment to perform analysis using the knowledge about the target, thereby easily obtaining analysis results based on industry knowledge and specialized knowledge.

[0056] The information processing system 1 of the embodiment further includes an AI model learning unit 24 and an accuracy evaluation unit 21. The AI ​​model learning unit 24 causes the AI ​​model to learn learning data. The accuracy evaluation unit 21 calculates the accuracy of predictions made by the AI ​​model. The AI ​​model learning unit 24 performs an improvement process in which the AI ​​model learns learning data that reflects the analysis results (XAI report) generated by the report generation unit 13. The accuracy evaluation unit 21 calculates the prediction accuracy of the AI ​​model that has been improved by the AI ​​model learning unit 24. The prompt generation unit 12 generates a prompt for causing the generated AI to analyze the prediction basis of the improved AI model, depending on the prediction accuracy of the improved AI model. As a result, the information processing system 1 of the embodiment can evaluate the AI ​​model that reflects the improvement measures, and analyze the prediction basis based on the evaluation results.

[0057] In the above-described embodiment, an example was described in which an analyst M considers improvement measures for an AI model. However, this is not limited to this. The XAI report generation device 10, or a computer different from the XAI report generation device 10, may be configured to consider improvement measures based on the XAI report. The following describes an example in which a functional unit (referred to as an improvement measure consideration unit) of the XAI report generation device 10 considers ways to improve the accuracy of predictions made by the AI ​​model.

[0058] The XAI report generation device 10 includes an improvement measure examination unit (not shown). The prompt generation unit 12 generates a prompt that instructs the generation of an XAI report including a proposed improvement measure. The improvement measure examination unit executes an improvement measure according to the proposed improvement measure indicated in the XAI report. For example, if the XAI report proposes the generation of training data to improve prediction accuracy as an improvement measure, the improvement measure examination unit generates training data to improve prediction accuracy according to the improvement measure. Alternatively, if the XAI report proposes reviewing the clustering of training data as an improvement measure, the improvement measure examination unit reclassifies the training data according to the improvement measure. Alternatively, if the XAI report proposes noise removal and / or mislabeling correction as an improvement measure, the improvement measure examination unit removes noise and / or mislabeling according to the improvement measure.

[0059] The information processing system 1 and the XAI report generation device 10 in the above-described embodiment may be implemented in whole or in part by a computer. In this case, a program for implementing the functions may be recorded on a computer-readable recording medium, and the program may be loaded and executed by a computer system. Note that the term "computer system" as used herein includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and CD-ROMs, as well as storage devices such as hard disks built into a computer system. Furthermore, the term "computer-readable recording medium" may also include devices that dynamically store programs for a short period of time, such as communication lines used when transmitting programs via networks such as the Internet or telephone lines, or devices that store programs for a fixed period of time, such as volatile memory within a computer system serving as a server or client. The program may also be designed to implement some of the functions described above, or may be capable of implementing the functions in combination with a program already stored in the computer system, or may be implemented using a programmable logic device such as an FPGA.

[0060] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and includes designs within the scope of the gist of the present invention. [Explanation of symbols]

[0061] 1. Information processing system 10...XAI report generator 11...Prediction Basis Acquisition Department 12...Prompt generation section 13...Report generation section 14...Report display section (report output section) 20...Learning device 21...Accuracy evaluation section 22...Prediction basis generation unit 23...AI model memory section 24…AI model learning section 25...Learning data storage unit

Claims

1. A prediction basis acquisition unit that acquires prediction basis indicating the basis of prediction by the AI ​​model; a prompt generation unit that generates a prompt to cause a generation AI to generate an analysis result corresponding to the prediction basis from the prediction basis acquired by the prediction basis acquisition unit; a report generation unit that generates an analysis result of the prediction basis by inputting the prompt generated by the prompt generation unit into a generation AI; a report output unit that outputs the analysis results generated by the report generation unit; An information processing system comprising:

2. The prediction basis acquisition unit acquires, as the prediction basis, the contribution of the feature used by the AI ​​model for prediction using SHAP (SHapley Additive exPlanation). The information processing system according to claim 1 .

3. The AI ​​model is a model that makes predictions using images, The prediction basis acquisition unit acquires, as the prediction basis, information indicating an area of ​​the image focused on by the AI ​​model in the prediction using GradCam (Gradient-weighted Class Activation Mapping). The information processing system according to claim 1 .

4. The prediction basis acquisition unit acquires, as the prediction basis, information indicating a degree of influence that affected the accuracy of prediction in the learning data learned by the AI ​​model, using an Influence Function. The information processing system according to claim 1 .

5. The prompt generation unit generates the prompt including knowledge about a target to be predicted by the AI ​​model. The information processing system according to claim 1 .

6. An AI model learning unit that causes the AI ​​model to learn learning data; an accuracy evaluation unit that calculates the prediction accuracy of the AI ​​model; further comprising: The AI ​​model learning unit performs an improvement process in which the AI ​​model learns learning data that reflects the analysis result generated by the report generation unit, The accuracy evaluation unit calculates the prediction accuracy of the AI ​​model that has been improved by the AI ​​model learning unit, The prompt generation unit generates the prompt for analyzing the prediction basis in the AI ​​model on which the improvement process has been performed, according to the prediction accuracy of the AI ​​model on which the improvement process has been performed. The information processing system according to claim 1 .

7. An information processing method performed by a computer, comprising: The prediction basis acquisition unit acquires prediction basis indicating the basis for the prediction by the AI ​​model, a prompt generation unit generates a prompt from the prediction basis acquired by the prediction basis acquisition unit to cause a generation AI to generate an analysis result corresponding to the prediction basis; a report generation unit inputting the prompt generated by the prompt generation unit into a generation AI to generate an analysis result of the prediction basis; a report output unit that outputs the analysis results generated by the report generation unit; Information processing methods.

8. On the computer, Obtaining prediction grounds that indicate the grounds for predictions made by the AI ​​model; generating a prompt from the prediction basis to cause a generation AI to generate an analysis result corresponding to the prediction basis; The generated prompt is input to a generation AI to generate an analysis result of the prediction basis; outputting the analysis results; program.

Images