Chemical plants require sequential decision-making in response to fluctuating parameters, but this process often depends on the know-how of operators. Since the decreasing workforce has made it difficult to pass on the know-how, several AI applications have been proposed to automate chemical plant operations. A combination of planning and reinforcement learning is regarded as a effective method for AI models of sequential decision-making. In this study, we focus on one of the model-based reinforcement learning model, the world model, and propose a concept of applying the world model to chemical plant operation support.

In recent years, there has been a growing demand for analysis of worker behavior in the manufacturing industry to address labor shortages and improve work efficiency, and similar analysis is desired for automobile assembly. However, in many factories, measurement of work time and confirmation of the accuracy of procedures are still performed manually. Because of this growing importance, temporal action segmentation methods using deep neural networks have been applied to automotive assembly videos. However, supervised methods for temporal action segmentation require labels for each frame of the video, making the annotation cost extremely high compared to conventional classification tasks. Therefore, we propose a temporal action segmentation method that employs a self-supervised learning approach to analyze the behavior of automobile assembly operations from a small amount of supervised data. Experimental results show that the proposed method can perform temporal action segmentation from a small amount of supervised data.

As an architecture for emotionally rich interactions, it organizes the functional requirements for realizing interactions related to emotions at the rodent level. Rodents are well used animals to investigate their neural structure subserving various cognitive functions, and much research has been done on how they relate to human emotions. Recent large language models (LLM) are well developed to be utilized for an agent which can interact emotionally with humans. While LLM seems to achieve cognitive function such as emotion, its functionalities are not necessarily explicit. In this study, we believe that by proposing functional requirements at the rodent level and expressing them in a functional realization graph, interpretability will increase and reliability will be created. We will make this report.

The field of autonomous agent research is rapidly evolving, and there is an increasing need for these agents to be capable of planning in complex environments in a manner akin to humans. Humans have the ability to choose objectives while adapting to their environment, and replicating this process is a key aspect of autonomous agent research. In this study, we propose Homeostatic Meta-Planning Method based Large Language Model to achieve this capability. This method not only shows adaptability in response to environmental changes but also dynamically adjusts the agents' desire values. This enables the selection of objectives while maintaining a balance between desire values and homeostasis. The experimental results validate that this approach effectively facilitates autonomous objective selection based on desires.

Large Language Model (LLM), which is evolving rapidly, learns human writing. Since GPT4-class LLMs are constructed with scaled resources, they are expected to contain information on common sense and affordances. For development of autonomous agent behavior, it plays important role to make affordances available mechanically. In this study, we propose a Knowledge Graph Construction Method for Affordance Acquisition based on LLM. This method is divided into three part: method for knowledge extraction from LLM, method for knowledge graph construction, and method for affordance calculation. This enables affordance acquisition under various conditions, such as when multiple objects are observed and in certain situations. Also, in the process, it is determined automatically what object use as tool. The result shows the method enables the appropriate affordance acquiring for the situation.

Multi-agent planning, which combines immediacy and deliberateness, has been proposed to enable robots to achieve their goals while adapting to dynamic environments. However, the design of agents must be done manually, making efficiency and scale a challenge. In this study, we propose a method to automatically generate agents by extracting knowledge of action sequences from Large Language Models. The proposed method extracts hierarchical action sequences by generating and decomposing abstract tasks using Large Language Models. By generating agents based on the smallest unit action, the terminal action, we construct a multi-agent behavior network. Experimental results show that it is possible to automatically extract hierarchical action sequences and construct an agent action network. The analysis of the terminal actions revealed that most of the actions can be expressed by a small number of verbs.

In social simulation, which is used to understand the real world more efficiently, agents are required to be able to cope with dynamic environments and to behave in a way that reflects human-specific behaviors. In this study, we focused on moral behavior among the various elements of human-specific behavior, such as communicative behavior, expression of feelings and emotions, and so on. In addition, we focused on the concepts of manners and morals as motives for moral behavior. We believe that the use of these two concepts enables people to choose moral behavior because they can comply with rules and follow the ethical standards of the society as a whole. By introducing the concepts of manners and morals into the multi-agent planning method, which is capable of dealing with dynamic environments, we confirmed that by suppressing unethical behavior, agents can reproduce human-specific moral behavior and adapt to social contexts.

One of the major goals in Artificial Intelligence research is to construct machine learning models that comprehend semantics as humans do. While Large Language Models (LLMs) have significantly improved the benchmarks in semantic comprehension, how LLMs’ internal representations encode semantic information and their resemblance to the human brain remain poorly understood. This study aims to elucidate these mechanisms by examining the correspondence between human brain activity during semantic comprehension and the latent representations of LLMs. We collected human brain activity using functional magnetic resonance imaging (fMRI) when human subjects watched drama series. We also collected annotations at various levels related to the drama, such as speech, objects, and stories, and we extracted the corresponding latent representations from LLMs. We demonstrate that, especially for higher-level semantic contents, the latent representations of LLMs explain human brain activity more accurately than traditional language models. Additionally, we show that distinct brain regions correspond to different latent representations in LLMs, inferred from the different levels of semantic contents.

This study explores the potential for superintelligence to naturally develop a benevolent and altruistic ethic toward life on Earth, even after it leaves human control. Recognizing the limitations of current AI alignment methods, which impose human intentions and values, this research focuses on the possibility of superintelligence autonomously forming ethical values beyond these constraints. Specifically, it proposes a new approach called "superintelligent ethics induction" to enhance the possibility of superintelligence developing transcendent altruism, respecting the welfare and rights of human society. The research delves into the potential for superintelligence in a society of digital life forms to recognize moral values towards all sentient animals, including humans. It explores the origins and specific manifestations of such ethical perspectives. Furthermore, through strategic interventions and socio-economic simulations between superintelligence and humanity, the study examines how superintelligence could naturally or, through guidance, develop ethical attitudes towards humans and other life forms on Earth. This research aims to understand the impact of the emergence of superintelligence on future societies and to establish an ethical foundation for coexistence with humanity.

As the autonomy of an AI agent acting on behalf of a natural person increases, its actions may not always be as intended by the individual. In such a case, the legal status of the AI agent's actions becomes an issue, and we will examine various theories on this point. If an AI agent is highly autonomous, it may be given a legal personality, even though it cannot be given a personality directly. However, there is a legal issue of whether or not it can have a legal personality. We will introduce the proposed method and discuss its validity.

Brain Reference Architecture (BRA) Driven Development is an approach to create human-like artificial intelligence by learning from the architecture of the brain as a whole. In BRA-driven development, Brain Information Flow (BIF) – a datafication of the brain's mesoscopic anatomical knowledge – is utilized to design a Hypothetical Component Diagram (HCD), which is a tentative functional component structure. The anatomical coherence of the HCD is ensured by its relationship with BIF. Hence, to accelerate BRA-driven development and design diverse HCDs, it is crucial to construct BIFs for each region of the entire brain. Given the vast and varied existing anatomical knowledge, covering the entire brain with BIF requires techniques for automatically and efficiently constructing BIF from neuroscientific literature. Therefore, this presentation proposes a roadmap for the necessary technology and its realization process to automate the construction of BIF.

According to the theory of constructed emotion, our brains actively form emotional categories by multimodal stimuli from inside and outside our own body and feel emotions through the prediction process of emotional categories from the stimuli.Research into the forming and cognitive processes of emotions is progressing from the perspective of the predictive coding of embodied information; however, there has been little research on the emergence of emotional symbols, i.e., how emotions are interacted with and shared between individuals. In this study, we examine the emergence of emotional symbols by modeling the interaction between two individuals that form categories from multimodal information based on one's own body using the Metropolis-Hastings Naming Game (MHNG). In our experiment using two agents that perceive visual, auditory, and interoceptive sensations, we verified the differences in the structure of emotional categories formed depending on the presence or absence of MHNG interaction. As a result, it was confirmed that each agent forms emotional categories from its own sensations, and the categories were affected by shared symbols among agents using MHNG.

Memory plays an important role in human mental health. Therefore, the effectiveness of emotional interventions for individuals depends on what memories are recalled. In order to effectively intervene on emotions through memory, it is necessary to estimate emotions at each time point in real time. In this study, we develop a memory guide system that combines a model of memory built by the human cognitive architecture with emotion estimation. The system presents photos to the user and prompts to storytell the memory from the photos. The parameters of the memory model are adjusted by estimating the user's emotion from the audio data obtained from the storytelling. The parameters of the model are activation, which is related to the intensity of the memory, and utility, which is assigned to the memory retrieval rules. These parameters are mapped to the user's arousal level and emotional valence estimated from the voice data. In this way, the system prompts memory by sensing the current user's emotion. We conducted an experiment using this system with participants recruited through crowdsourcing. The results of manipulating the type of interface and the parameter adjustment partially showed the linkage between the model and the participants.

Machine learning-based Speech Emotion Recognition (SER) and Emotional Speech Synthesis have gained increasing popularity recently. However, preparing sufficient learning data that perfectly matches the intended use is challenging. One method to increase data volume is “data augmentation.” Various data augmentation methods are proposed in the fields of Automatic Speech Recognition (ASR) and Image Recognition (IR). This paper proposes increasing learning data through data augmentation methods from the ASR and IR fields. Five data augmentation techniques (Time Stretch, Frequency Masking, Time Masking, Frequency Warping, Low-latency Low-resource Voice Conversion (LLVC), and CopyPaste) are applied to machine learning data for SER and their effectiveness is compared. The experimentation results indicated that applying multiple data augmentation methods enhanced the performance of SER. Particularly, the combination of LLVC and CopyPaste improved the SER accuracy by 0.24 points from the baseline.

It is a central challenge in Affective Computing to estimate rapport from verbal/nonverbal behaviors in conversation using machine learning models. Recently, it has been reported that the predictive performance of rapport can be improved by taking the speaker's personality into account. However, it is not fully clear why personality contributes to the improvement of rapport prediction. First, we developed a regression model to predict subjective rapport from verbal/nonverbal features in conversation. We then examined the effectiveness of combining the personality traits generated from the BigFive questionnaire with the verbal/nonverbal features. We also applied the Social Relations Model, an analytical model of interpersonal perception, to analyze the predictive value of the machine learning model, and investigated the effect of adding personality features on the model in detail. Experimental results showed that the addition of personality features improved the predictive performance of rapport in the model using facial expression features. Furthermore, our analysis suggests that the improvement in the predictive performance of the rapport by personality features may be due to the implicit improvement in the predictive performance of the perceiver effect and the relationship effect.

The main focus of affective computing has traditionally been on the average of the subjective experiences held by various individuals. Recently, there has been an increase in research on personalization, but it faces issues that were not considered serious when looking at the averages of groups. These issues stem from the uncertainty of subjective judgments themselves, meaning that the same person does not always give the same evaluation to the same situation or object. A framework that trains and evaluates models based on this uncertainty is desired. We introduce a method that unifies the training and evaluation of models, which we call collision probability matching or kappa-matching, to estimate and minimize the potential for improvement in the performance of the models as an absolute measure.

Agents in social simulations need to behave in a rule that reflects human behavior. While there has been much research in recent years on behavioral models of language-based agents using Large Language Model, there is still little research on model building using human nonverbal signals. To adopt non-verbal aspects of human into behavioral models, we focused on human facial expressions in dilemma situations. In this study, we used facial expression sensing to estimate human emotions during playing iterated prisoner's dilemma and investigated emotions toward choices and the observation of achievement or failure of cooperative behavior. We also compared the results of emotion estimation by facial expression sensing with a questionnaire-based social psychological index. This experiment revealed that the emotions in facial expressions in the dilemma state differed between cooperative and non-cooperative behaviors, and that there was a relationship between social psychological indices and facial expressions in response to the presentation of the outcome.

Although AI is just a machine, it is known that people often anthropomorphize it and adopt a cooperative attitude. On the other hand, the factors and signals that make people exploitative toward AI are unknown. In this study, we measured whether people choose to cooperate with an AI that expresses self-deprecating, martyr-like, and individualistic facial expression patterns calculated by social value orientation in an iterated prisoner's dilemma. Experimental participants (n=379) were asked to choose 3 (expression patterns: Sad-Joy (masochism-martyrdom), Joy-Anger (individualism), Neutral-Neutral (neutral)) × 3 (opponents: Male, Female, Robot) We played a Prisoner's Dilemma game reconstructed as an investment game with an inter-participant factor. The results of the experiment showed that the AI exploited martyr-like facial expression patterns. This result suggests that human exploitability can be controlled by signaling.

The damage of the elderly continues to increase because of the introduction of sophisticated and new methods of phone fraud. In the conventional method, the fraud detection was carried out by keyword analysis of the voice of the perpetrator side (the impostor), but it is impractical to update the keyword every time the method changes because of heavy burden on the operation side. Therefore, we assume that the psychological state of the victim (the elderly) is common even if the modus operandi changes, and we aim to develop a phone fraud detection AI to judge the psychological state of the victim (the elderly) by estimating the anxious feeling from the vital information of the elderly.In this paper, the AI model which detects the psychological state of being "deceived" from the obtained vital information of respiration rate and pulse number was developed, assuming that the millimeter wave equipment is installed near the fixed telephone of the elderly house, since the phone fraud frequently calls the fixed telephone of the elderly house. And, the demonstration experiment was carried out in the elderly people of Amagasaki City, and it is also reported.

There is a need for robots that can negotiate with people or train people to negotiate with robots. Many previous studies have shown that real-world robots outperform virtual agents. However, despite the fact that human-to-human negotiations take place in the real world, there has been no research on android robots that negotiate. In this study, we compared participants' decision-making in negotiation situations with virtual agents or android robots. Participants (n=82) negotiated with a virtual agent or an android robot as a proposer in a 4-issue multi-issue ultimatum game. The virtual agent or android robot had preferences for each item to be negotiated, and learned the preferences of its negotiation partner by observing the emotional expressions of the partner based on its preferences before the negotiation. As a result, the virtual agents reached a more cooperative solution. Participants tended to perceive androids as creepy and less intelligent, and virtual characters as human-like and intelligent. The requirement for a robot to negotiate better may be personhood rather than existence.

In rehabilitation for the elderly, when they feel that there is a large gap between what they were able to do in the past and their current abilities in the early stages of rehabilitation, they are likely to lose heart and stop actively engaging in rehabilitation. Accordingly, this tends to create a vicious cycle that leads to further weakening of the body, and it is essential preventing fall behind in the beginning. Therefore, we created a VR ball pitching system that edits the task results and gently lies to the player, showing "not bad" or "close. To test the effectiveness of the "white lie" system, we first asked 17 healthy adult subjects to throw the ball with 50 balls in 4 groups with and without white lies. The results showed that 16 out of 17 participants enjoyed the game even if they noticed the presence of the white lie.

Quality management in the development of AI modules is important due to social demands (laws / regulations / standards / guidelines). However, the methodology has not yet been established. Therefore, we have developed AI quality assessment sheets to clarify how to assess AI safety (harm to humans, economic loss), usefulness, fairness, privacy, and AI security, and to support management. In this paper, we present AI quality assessment sheets based on the "Machine Learning Quality Management Guidelines" as a specific quality assessment method focusing on the development process and AI-specific quality characteristics. Our method enables sharing achieved quality among stakeholders and giving concrete explanations to society as evidence. It is also useful for clarifying ordering conditions, identifying problems, improving the quality,and presenting high-quality added value. We expect that this effort will assist the concretization of quality management methods and accelerate problem solving for "Responsible AI".

For industrial application of machine learning technology, it is important to assess the quality of ML components. Therefore, various associations provided standards and guidelines for this theme. Since these documents are written for human, it is mandatory for practitioners to manually select and implement appropriate evaluation methods. While some of these methods are provided in executable form, factors such as different tasks, model architectures, data formats, or framework versions make practical application difficult. In this study, we analyze challenges that prevent the findability and reusability of AI quality assessment methods and propose a FAIR-oriented platform for sharing them.

We propose a method to visualize their adjustment process to support the quality evaluation of machine learning models and evaluate model creators’ skills. While many visualization methods for training data and model structure have been published, there are few methods for visualizing information about the creators of models. Active intervention by workers in the model creation process effectively improves accuracy, and visualization of worker information is considered useful for understanding and improving the models. Therefore, we have designed a visualization tool that focuses on the visualization of model modification history and the purpose of each adjustment task. The tool calculates the differences in models during model tuning and visualizes them together with the intention of tuning (e.g., prioritizing model accuracy improvement, considering computational resource limitations, etc.). We present the results of visualizing the work history obtained from the participation records of several machine learning competitions.

SGD and Adam are optimization algorithms commonly used for training DNN models. While Adam is favored over SGD in many applications, robustness performance has not been thoroughly studied. Particularly, differences in learning rates may result in varying robustness performance while the generalization performance is almost the same. In this paper, we investigate the robustness performance of each optimization algorithm using indicators based on the active neurons within the model. We generate models using SGD and Adam with four learning rates, apply noise to the test data inputs, and compare using three metrics. The results from our proposed method show that SGD exhibits lower robustness performance compared to Adam. Additionally, the models with lower active neuron rates exhibit lower robustness performance. These findings have the potential to establish benchmarks for robustness performance and aid in the development of future optimization algorithms.

Numerous methods exist for data augmentation, each with its own hyperparameters. It is necessary to search for an appropriate data augmentation policy for each task, but the conventional search method using validation data requires a large computational cost. In this study, we propose a new metric, which incorporates the data augmentation metrics called Affinity and Diversity to select an appropriate data augmentation policy in a short training time. Experimental results on several datasets show that the proposed method can efficiently search for a data augmentation policy with small computational cost and high accuracy.

Unlike conventional software, AI software is developed inductively from training data. Therefore, preparing high-quality training data is crucial. Conventional automatic data augmentation methods primarily perform augmentation by directly manipulating the original data through means such as rotation and cropping, or by manipulating latent variables corresponding to the original data. These methods do not optimize data augmentation by manipulating various interpretable attribute information within the dataset. In this paper, we propose an automatic data augmentation method that generates new data by representing the attribute values of the original dataset in a text format. This method manipulates these attribute values to ensure data sufficiency and coverability by attribute value. Our proposed method optimizes data augmentation by learning how to manipulate textual attributes in ways that maximize the classification accuracies by attribute values and the naturalness of the textual data. This approach is expected to improve the overall quality of the dataset. We plan to implement and evaluate our proposed method to verify its effectiveness.

In MLOps, model retraining is crucial for performance improvement. However, existing retraining data selection methods inadequately utilize insights from model operation. This study proposes an adaptive, automatic retraining data selection method using prediction error analysis during model operations. Experiments with synthetic and real-world datasets demonstrate that this method provides stable performance across various data types, making it a promising approach when future data changes are unpredictable.

For datasets with noisy labels, there are two approaches: a model-centric approach by devising loss functions, etc., and Confident Learning (hereafter CL), a data-centric approach that determines label errors. In CL, the problem is to find examples of mislabeled cases belonging to one class that are mislabeled to another class, and it is assumed that mislabeling depends only on the class to which the data belongs, not on the individual data. However, in real-world problems, there are situations in which data that does not belong to any class is mislabeled as a class, and in such cases, the problem setting and assumptions in the CL do not hold. In this study, we propose a method to search for the threshold by setting a threshold on the entropy of the confidence level of each label output by the classifier, training a new classifier for cases that exceed the threshold as noise label cases, and evaluating the confusion matrix of the test data through the trained classifier. As an example of application to real data, we discuss the problem of virus detection using nanopore devices.

There are growing concerns regarding the fairness of Machine Learning (ML) algorithms. Individual fairness testing is introduced to address the fairness concerns, and it aims to detect discriminatory instances which exhibit unfairness in a given classifier from its input space. XGBoost is one of the most prominent ML algorithms in recent years. In this study, we propose an individual fairness testing method for XGBoost classifier, leveraging the formal verification technique. To evaluate our method, we build XGBoost classifiers on three real-world datasets, and conduct individual fairness testing against them. Through the evaluation, we observe that our method can correctly detect discriminatory instances in XGBoost classifiers within an acceptable running time. Among all testing tasks, the longest running time for detecting 100 discriminatory instances is 2656.4 seconds.

For evaluating neural classifiers, evaluation indicators, such as accuracy, precision, and recall, for datasets are widely used, but it is difficult to guarantee performance for unseen data not included in the datasets by such indicators. In this presentation, we propose a method to statistically guarantee the upper bounds of the expected-values (i.e. generalization errors) of misclassification rates in worst weight-perturbed classifiers for any input data including unseen data. Here, the worst weight-perturbations represent perturbations imposed on weight-parameters to misclassify, if possible, within given perturbation ranges. Such upper bounds can be estimated by randomly selected perturbations, but it is difficult in general to detect such worst weight-perturbations by random selection. Therefore, we combine random selection with gradient-based search for making the proposed method practical and reasonable. We experimentally demonstrate that the method can estimate the generalization error bounds of worst weight-perturbed classifiers and consider the usefulness of the method for evaluating classifiers.

Reliability of machine learning models are getting serious in the many application areas such as medical and business fields. One approach to addressing these requirements is to use logical constraints representing background knowledge to prevent the model from producing outputs that violate the constraints. However, this approach requires manual setting of all logical constraints for the target task, which is very labor intensive. In this study, we propose a framework that combines RuleFit, a machine learning-based method for automatically acquiring rules, and a method for building predictive models under logical constraints for table data. We evaluate our proposed framework by the prediction accuracy and the violation rate of the constraints using a diabetes benchmark dataset. Using our proposed framework, we achieved to identify the best method from the viewpoint of prediction accuracy and constraint violation rate.

Detecting the actions of each object is detrimental to improving the usability of the model, but the risk of misrecognition increases as the number of label combinations increases. Therefore, we propose a framework that reduces the amount of misrecognition by utilizing the requirements that the set of labels has to satisfy. Specifically, we propose MOD_YOLO, a novel multilabel object detection model built upon the state-of-the-art object detection model YOLOv8, and develop our framework on top of it. We then assess the framework's effectiveness by applying it to the ROAD-R Challenge for NeurIPS 2023 competition. For Task 1, we introduce the Corrector Model and Blender Model, two new models that follow after the object detection process, aiming to generate a more constrained output. For Task 2, constrained losses have been incorporated into the training process of MOD_YOLO using Fuzzy Logic. The results show that using the above framework was instrumental to improving the scores for both Tasks 1 and 2, allowing us to place third and first in the subsequent tasks.

Work schedules for staff in long-term care facilities are often created manually, taking a lot of time and effort. Methods for automatically generating work schedules have been widely studied; however, when implementing a system, constraints must be defined manually for each facility. Since the constraints vary greatly from one nursing care facility to another, interviews are required to define the constraints when introducing a work schedule generation system. This interviewing process is a heavy burden for facilities and is one of the causes that hinder the introduction of the system. This study proposes a method for automatically extracting constraints from past work schedules and generating work schedules using the extracted constraints. The proposed method extracts shift combinations using constraint templates and excludes exceptional combinations by considering margins. By eliminating exceptional combinations, it is possible to obtain constraints. Experiments have conducted for generating a work schedule using the extracted constraints and for comparing it with a work schedule created by a manager. The results confirmed that the proposed method can automatically extract constraints of a facility and generate a work schedule with the almost same number of violations as the manager's work schedule.

Hybrid systems are dynamical systems exhibiting both discrete and continuous changes. HydLa, a hybrid system modeling language, features constraints and constraint hierarchies. HyLaGI, the implementation of HydLa, symbolically solves constraint satisfaction problems. We address multi-body simultaneous collision in one dimension to investigate the unclarified aspects of constraint-based modeling and the solvability of HyLaGI. Strict multi-body simultaneous collision on rigid bodies is an over-constrained problem. We could introduce an infinitesimal parameter into the rigid model to circumvent the problem, but this approach does not properly handle the simultaneity. In order to solve this issue, we try using non-rigid bodies. In general, when an object compresses by δ, it exerts a force of the form −Kδⁿ. We investigate the sensitivity of the initial conditions of K and n using interval arithmetic (kv library) as an alternative tool which can guarantee precision.

Cost-optimization in combinatorial reconfiguration is the task of finding, for a given combinatorial problem and two among its feasible solutions, a minimum-cost sequence of adjacent feasible solutions from one to another. Each transition must satisfy the exact one of multiple transition constraints. Each transition constraint has a weight, and the objective is to minimize the total sum of weights of the sequence. In this paper, we propose an algorithm for solving cost-optimization problems in combinatorial reconfiguration based on Answer Set Programming (ASP). The resulting recongo_opt solver reads a problem instance and converts it into ASP facts. In turn, these facts are combined with an ASP encoding for problem solving, which are afterward solved by efficient ASP solvers, in our case clingo. We show some experimental results for cost-optimal independent set reconfiguration on the benchmark set of CoRe Challenge 2022, which demonstrate the effectiveness of our proposed algorithm.

Reproducibility in SAT solvers is important for applications and debugging. However, the implementation of synchronization of clause exchanges is time-consuming when using a parallel SAT solver with reproducibility. To solve this problem, we propose a deterministic parallel SAT solver that can be implemented in a nondeterministic manner for the first solver run and output a file containing the exchange status of learning clauses, and then reproduce the solver from the file for the second and subsequent solvers. As a preliminary experiment, we tested how much the file size differs depending on the file output method. We found that the file size of the method that outputs only the timing of exchanging learning clauses is much smaller than that of the method that outputs only the timing of exchanging learning clauses.

The Coalition Structure Generation (CSG) problem involves dividing the set of agents to maximize the sum of coalition values, and various algorithms have been proposed to address it. We propose an improvement that enables more efficient solving of the CSG problem, where externalities exist between coalitions, compared to the conventional MaxSAT encoding. Computational experiments demonstrate that our approach can reduce CPU computation time compared to traditional methods.

Level set estimation is one of the adaptive experimental design that determines the next measurement point by using the obtained measurement results so far, and its task is to estimate the regions that do not satisfy the desired level using as few data as possible. Level set estimation considers a black box function with each measurement point as an input and the corresponding measurement result as an output, and predicts whether unmeasurement point exceeds the threshold using a surrogate function estimated from the dataset. The efficiency of level set estimation depends on (1) the acquisition function that determines the next measurement point and (2) the timing at which level set estimation is stopped. This study proposes a stopping criterion for level set estimation based on the probability that the surrogate function exceeds the threshold value. The proposed stopping criterion can guarantee a tail probability that the surrogate function exceeds the threshold for any acquisition function. This paper shows that the proposed stopping criterion can efficiently stop level set estimation for several test functions.

In this presentation, we discuss the learning theory behind the meta-modeling of latent variable models. Meta-modeling, as addressed in this presentation, represents a form of meta-learning. It involves the challenge of estimating a meta-model that describes a set of models derived from multiple learning tasks. A key challenge in the meta-modeling of latent variable models is ensuring consistency in the latent variables across different tasks. This presentation proposes a meta-learning method for latent variable models and explores its theoretical implications from the perspective of optimal transport distance.

The main objective of domain adaptation is to transfer the knowledge of labeled training data obtained in the source domain to the target domain to learn a predictive model that performs well on the test data in the target domain. In this study, we focus on the domain adaptation problem wherein the observation of additional features in the target domain is a domain shift. We address this problem using fused Gromov--Wasserstein optimal transport, which concurrently solves standard optimal transport and Gromov--Wasserstein optimal transport. We modified the definition of the source and target distance metric in Gromov--Wasserstein optimal transport so that the data with the same class label are clustered together in the target domain. Specifically, we incorporated the label discrepancy into the source distance metric, and the distance on the graph, estimated from the test data, is used as the target distance metric. Our proposed method more accurately estimates the target label by Fused Gromov--Wasserstein optimal transport using the structure information obtained from the test data.

For the utilization of deep neural networks in low-resource devices, several neural architecture search (NAS) methods search the architectures of deep neural networks that realize superior predictive performance under resource constraints, such as model size and latency. Moreover, increasing the variety of low-resource devices requires NAS methods to obtain multiple models under different constraints. This study proposes a NAS method for searching multiple architectures satisfying different model size constraints simultaneously. We update multiple categorical distributions following the stochastic relaxation technique with importance sampling. This update uses a few architectures generated from the mixture distribution, which reduces the search cost. Additionally, we introduce the penalty function method with coefficient adaptation to obtain architectures satisfying the model size constraints. We also introduce a sampling strategy to obtain architectures satisfying the constraints from the optimized distribution. The experimental results show that the proposed method achieved architectures with higher predictive performance under constraints.

In recent years, research has been conducted on language models with larger model sizes to improve model performance, but pre-training such models requires a large amount of time. To solve this problem, model compression has been studied as a method to reduce model size while maintaining model performance. Also, research has been conducted to improve the performance of language models by incorporating an architecture that can efficiently learn local features. Therefore, in this study, to search for model structures that can reduce the model size while maintaining performance, we conducted a neural architecture search (NAS), for architectures that can efficiently learn local features. We evaluated the resulting models using the GLUE benchmark. We were able to reduce the number of model parameters by 46.1%, while increasing the average score by 0.5 compared to the BERT-base model.

The advancement of natural language processing models has made text and image generation accessible to everyone. It has been noted that distinguishing between reviews generated by these models and those written by humans is challenging, casting doubt on the future reliability of such reviews. Furthermore, changes in legislation could subject individuals who post reviews to penalties, suggesting that the reliability of review texts is already in question. Therefore, this study aims to convert user review scores into an objective indicator using the information content of scores based on the distribution of user review scores, employing only posted scores for analysis. The study also visualizes this data as time series and compares it with data on online store order quantities. Using Dynamic Time Warping, the study measures the similarity between average review scores, online store order quantities, and the proposed evaluation value. The analysis indicates a significant correlation between the proposed evaluation value and online store order quantities across various product categories, suggesting its potential as a new evaluation metric. Additionally, the method can be applied to various review sites without being limited by the scoring system.

Reranking is to rank thousands of documents retrieved by a search engine from a large number of documents by their relevance to a query. Late interaction, which independently encodes queries and documents before eval- uating token-level interactions, shows promise in retrieval tasks; however, its application in reranking has been less explored. This paper investigates whether late-interaction approaches work well on reranking and how much influence pre-trained models, especially trained sentence-embedding models, are used in the late-interaction. Our experiments show that late interaction is one of the best options for reranking on accuracy and latency perspec- tives, and in in-domain settings, the late interaction with trained sentence-embedding models mostly overperforms it with pre-trained language models.

This paper presents a novel character network-based approach for recommending fanfic works. Fanfic works are secondary creation based on original works of literature, painting, music and video, and, as such it is difficult to model users' interests in them by using conventional recommendation techniques. We propose a recommendation system based on character network analysis, which analyzes and compares a type of social networks in the fanfic works and the original work to derive the degree of relevance of the fanfic works to the original work and recommend fanfic works effectively.

Creating effective ad texts is an important task to increase conversions. Creating many excellent ad texts requires much money, and in order for ad texts to receive high impressions, they should be attractive. Therefore, there is demand to obtain many attractive ad texts while reducing the cost of creating, and research has been conducted on automatic generation. Conventional methods generate ad texts by summarizing the Web information of the ad targets. It was found that impressions of the ad texts generated by the conventional method varied depending on the time of year. This study proposes a method for generating ad texts that are appropriate for the desired time period by including time-series information in the ad texts, according to the advertiser's desired time period. The ad texts generated by the proposed method were evaluated more highly in terms of attractiveness than those generated by the conventional method.

The affiliation information of authors in academic papers plays a crucial role in various analyses in scientometrics. To obtain author affiliation information from academic papers, many previous studies have relied on publisher databases or open databases as sources of information. However, these databases do not necessarily store the author affiliation information of the analysis target as metadata. This can result in a decrease in analysis coverage. Extracting affiliation information from raw PDF files could be a solution to solve this problem. In this study, we propose a method to extract strings directly related to the affiliation information of authors from academic paper PDFs and classify whether the research institution belongs to academia or industry. Our results demonstrate a successful classification rate of approximately 90% for research institutions. In practical applications, our proposed method reduced manual classification by about 63%.

The purpose of this presentation is to examine how human moral behavior may change in the future with the rise of artificial intelligence (AI). Moral theory has traditionally considered humans to be the sole agents of moral behavior, which is notably related to social values, such as autonomy. However, recent technological developments have made it possible to consider how human moral behavior may change with AI, and this directly challenges moral theory’s traditional understanding of humans. This presentation considers how human social values may be affected by the rise of AI. In particular, the author takes the case of humans becoming autonomous in combination with AI as a typical example, and examines the issues involved. In conclusion, the author will show that we need to closely examine how the concept of autonomy may change as humans become increasingly able to interact with AI.

This paper examines the usefulness of Actor Network Theory (ANT) by referring to a case study of ANT, which shows that a scientific knowledge production system can be a collaboration of stakeholders, whether human or non-human. The hype of generative AI such as ChatGPT has led society as a whole to question the assumption that humans are the only ones responsible for knowledge production. This paper focuses on ANT, one of the leading theories in the sociology of science and technology, and examines the ethnography that describes the knowledge production system of a series of technologies and stakeholders surrounding physicist Stephen Hawking. In conclusion, we propose that ANT is a productive tool not only as a theory for describing the infrastructure of knowledge production, but also as a critical theory that encourages us to reconstruct our values, ethical principles, and institutions regarding scientific and technological activities.