Knowledge Tracing (KT), the task of tracing students’ knowledge state has attracted attention in the field of artificial intelligence. Recently, many researchers have proposed KT methods using deep learning to predict student performance on unknown tasks. Especially, the latest DeepIRT reportedly has high prediction accuracy and parameter interpretability. Nevertheless, some room remains for improvement of its prediction accuracy because it does not optimize the degree of forgetting of past data. Specifically, although its forgetting parameters are optimized solely using current input data, it should use both current input and past data to optimize the forgetting parameters. To resolve that difficulty, this study proposes a new DeepIRT with hyper-network that optimizes the degree of forgetting of past data using both the current and the past data. Results obtained from experimentation demonstrate that the proposed method improves the prediction accuracy and the interpretability of the students’ ability compared to earlier KT methods.

Adversarial Examples are malicious input data created by adding a small perturbation to original input data, and these data make a classifier outputs a wrong result. For object detectors, there exists a method to create Adversarial Examples Patches which can be printed and applied to object in real world, so detectors dismiss that object. In this work, we propose a defense method to detect if an Adversarial Examples Patches attack is taking place or not against an input image. Our method infers the position of Adversarial Examples Patches and paints those using bounding boxes with score values which are lowered below a threshold by an attack. We show our method is sufficiently better than random classification by evaluation over INRIA Person Dataset.

In this paper, we aim to improve the accuracy of solar radiation prediction used for photovoltaic power generation by constructing a classification model that captures the characteristics of weather data. In our previous studies, we used the weather conditions published by the Japan Meteorological Agency for classification. In this study, we use a clustering method, which is unsupervised learning, for weather classification to create a new weather situation classification. We aim to improve the accuracy by constructing a weather situation classification model that is more suitable for solar radiation prediction than before.

Electricity is difficult to store, so it is important to maintain a balance between supply and demand. There are conventional machine learning methods that use RNN, which can learn short-term dependencies, but have difficulty learning long-term dependencies. LSTM, a kind of derivative of RNN, is proposed as one of the solutions to this problem. In this paper, we aim to improve the accuracy of electricity demand forecasting by comparing and examining the effects of these two methods on electricity demand forecasting.

Entity Linking (EL), the task of mapping entity names in natural language sentences to resources in large-scale knowledge graphs, is attracting attention as a fundamental technology for question answering and dialogue systems, etc. DBpedia Spotlight (DS) proposed as an EL tool for DBpedia. Although DS supports multiple languages, the target language model of OpenNLP, a natural language processing library, is required to perform EL specific to a particular language. DS's multilingual model can be used for Japanese EL, but its accuracy is lower than that of OpenNLP's target language model. As of January 2022, both the Japanese model of OpenNLP and the Japanese model of DS have not been released. In this study, we aim to develop a Japanese model of DS by introducing the Japanese morphological analyzer Sudachi into DS. We showed the effectiveness of the Japanese model by comparative evaluating the model of DS and the multilingual model.

The “Knowledge Graph Reasoning Challenge” is a technical contest to organize and promote AI technology with explainability. The challenge consists of two tasks: 1) guessing the culprit and 2) explaining the reason for the guess, using a mystery novel represented as a knowledge graph. The knowledge graphs we provide have been improved by adding target novels and considering the contents to be described as knowledge graphs, keeping in mind that they are used for explanation generation. This paper proposes a guideline for constructing knowledge graphs for explanation generation, based on our previous studies. Based on the knowledge graphs of eight short mystery novels published in the Reasoning Challenge, we examined the revision policy of the contents of the knowledge graphs. We made a draft guideline for constructing knowledge graphs consisting of ten items. The guidelines are applied to the eight knowledge graphs, examining their validity. In this presentation, we report the outline of the guideline and the findings obtained by applying it to the knowledge graphs.

To process human knowledge with computers, various technologies are being studied, such as knowledge graphs. To enable sharing or reuse of knowledge by different applications, there must be rules or guidelines so that the same knowledge can be represented with the same knowledge graphs. The author proposed using the Entity-Relationship Model (ERM) for defining the concepts to be used in knowledge graphs. In this approach, individual knowledge graphs are constructed by instantiating an ERM, and knowledge graphs constructed this way can be shared or reused by different applications. However, the ERM can only represent simple propositions. This paper proposes using the Segmented Discourse Representation Theory (SDRT) for representing complex propositions as knowledge graphs. With this method, it is possible to construct complex knowledge graphs that can be shared or reused by different applications. This paper also proposes using FrameNet for defining the ERM so that the knowledge graphs constructed from the ERM can be shared or reused among different domains.

RDF data shows their values the most when they are built in a distributed manner and linked together from several aspects with URIs as the keys. However, we have seen several URI mismatch that should be identical from case discrepancies to misuse of symbols such as ‘#’ and ‘_’. Therefore, RDF curation is needed to make RDF data more valuable. We propose an infrastructure where RDF data developers can check, validate, and edit the data.

Knowledge graphs are knowledge representations that focus on relationships among objects. They have been used for question answering systems and information retrieval. As datasets become larger and leverage multi-modal representations, it becomes more important to complement missing or insufficient information in a single knowledge graph using other knowledge graphs with additional information such as images and attribute values. The entity alignment is a task of finding objects with the same object in different knowledge graphs, and multi-modal entity alignment (MMEA) has been proposed for the entity alignment of multi-modal knowledge graphs. However, MMEA does not take into account well the granularity of each piece of information since it represents each piece of information obtained from images, relations, and attribute values by a single point in a common space. In this study, we propose a new method that expresses the granularity of each piece of information as the spread of a distribution. The proposed method outperforms MMEA in the entity alignment task of two multimodal knowledge graphs.

We have been studying a model for generating facial expression videos that reflect the emotions of dialogue content in order to improve the human-like nature of dialogue agents. In a previous study, we proposed a model that can generate human-like facial expressions by learning the knowledge of lip-sync expressions and emotional facial expressions from different datasets. However, the generation results are inadequate due to the use of phonemes as input data and the frame rate of the generation results being too low. In this paper, we improve the model proposed in the previous study by using video as the input data and increasing the frame rate of the generated results to improve the quality of the results. In addition, by inputting the expression point video generated by the model into the facial expression video generation model for real images, we can generate facial images for emotional speech videos. For the facial expression video generation model, we use a model proposed by Zakharov et al. that can transfer facial expressions to arbitrary face images . The generated facial expression videos are subjected to sensitivity evaluation.

One way to improve the relationship between humans and anthropomorphic agents is to have humans empathize with the agents. In this study, we focused on the task between agents and humans. Therefore, we experimentally investigated hypotheses possibility that task difficulty and task content would facilitate human empathy. The experiment was a three-way multivariate analysis of variance (MANOVA) with eight conditions: task difficulty, task content, and empathy values before and after the task (2 x 2 x 2). The results showed that there was no main effect of the task content factor, and the main effect of the task difficulty factor tended to be significant. In addition, there was an interaction between task difficulty and empathy values before and after the task. As a result, when the task difficulty was high, empathy was facilitated more than when the task difficulty was low.

Active listening is a crucial attitude with positive acceptance and empathy in counseling. This study proposed and designed the active listening agent, which notified the internal state, using Artificial Subtle Expression (ASE). It is a method to notify an internal state considering versatility and easy introducing to various agents, such as light blinking and beep sound. We incorporated the function of synchronous response to negative emotion into ASE. The agent judged its emotion from speech volume recorded in real time and output ASE's increasing beep sound, which corresponds to the speaker's voice quality. These were referred by the findings that empathy is encouraged when an agent's facial expression matches a participant's emotional state and that in negative emotion, speech volume is louder and the feature is easy to judge. We primarily explain the basic structure of the designed active listening robot and ASE's algorithm.

When an operator teleoperates with semi-autonomous conversational robots simultaneously installed at multiple locations, the operator needs to quickly understand the context of the situation to change the operating robot. Therefore, clarifying suitable summarization methods for each typical situation for the use-case of multiple conversational robots provides sufficient conversational services. In this paper, we proposed a summarization method for each situation and researched if operators can understand the context more quickly by the proposed method than the standard method. Firstly, we categorized conversation situations by task-orientation of possible use-case in conversational robots, and then we focused on the "taking out of food" and "chitchat" as a task-oriented and non-task-oriented conversational situation. Finally, we researched the suited summarization method by comparing the proposed method and the common one. As a result, we found that the proposed method suits for focused conversational situations to understand the contents of a conversation.

In order for dialogue systems to continue to be used in everyday situations in the home, we believe it is effective for agents to have preferences and initiative and to have a position other than dialogue in everyday life. In this study, we aimed to build a collaborative music listening agent with the following functions: 1) to form preferences based on the songs it listened to; 2) to recommend songs based on the preferences of itself and its partner; 3) to play music for which preferences are unknown (curiosity-driven recommendation); 4) to select utterances for which the response of the partner is likely to be positive; 5) to try utterances for which the response of the partner is difficult to predict (curiosity-driven dialogue), and; 6) to control the dialogue based on the estimation of knowledge held by a large-scale language model. This paper discusses the future issues and prospects revealed by constructing a prototype version.

A learner can understand the complex and undetermined world through experiential learning. This study considers experiential learning as a process of symbol emergence derived from multimodal interactions between a learner and the real world. The purpose of this study is to construct the computational representation of a learner's internal information processing to generate experiences, and to develop a basic model for estimating the generation state of experience. A feature of our model is integrating a belief system (often implemented in robotics) with cognitive science. We analyzed multimodal data of environmental learning as a model case of experiential learning. Our qualitative analysis found that behaviors to update past experiences by behavior results and to generate experiences by belief. This study can be a basis for the next generation of HCI (Human-Computation Interaction) research, including the development of intelligent user interfaces that encourage people to change their behavior.

The market for content creation has been expanding, and the demand for stories and scenarios has been increasing. However, the shortage of scenario writers and the limitation of story patterns are serious problems. There is a need to develop a system to stimulate writers' creativity in scenarios generation. Since the direct automatic generation of a scenario is a hard task, we try to generate a plot, which is the main idea of a scenario, automatically. In this study, we propose three methods for automatic plot generation that use a narrative structure. We compared the plots manually generated with those generated automatically with our methods and using a narrative structure and confirmed the usefulness of our methods in automatically generating plots using a narrative structure.

Although the performance of task-oriented dialogue systems has been improving, not all users are able to accomplish their tasks perfectly. In particular, users with little knowledge of the system may not know how to effectively communicate with the system, resulting in dialogue breakdowns and failure to accomplish the task. In this study, we aim to improve the system's understanding of user utterances by providing a tutorial at the beginning of a dialogue to notify users of the type of utterances that the system can understand. We built a tutorial system using the MultiWOZ dataset and verified its effectiveness through utterance understanding experiments with human users.

When creators draw landscape paintings, they often refer to real buildings, but it takes time to find the building that they want. On the other hand, the generative adversarial network (GAN) is one of image generation technologies that have attracted attention in recent years. The Pix2Pix, a derivative technology of GAN, enables image-to-image translation. In this background, we think it is possible to generate design ideas by generating realistic images of buildings from conditional rough images using Pix2Pix, and to shorten the time for them to create new ideas. In this research, we developed a design support system that generates realistic building images in real-time by having system users draw label images. We evaluated usability by System Usability Scale and conducted a questionnaire survey.

Ryuteki is one of the instruments in the traditional Japanese music, Gagaku. It is difficult for beginners to learn its playing techniques. The best way to learn its playing techniques is to learn from a skilled person such as an instructor, but Gagaku has limited opportunities to learn from skilled person, especially due to time and cost constraints. Therefore, in this research, we have developed a system that automatically determines whether the pitch and timing of the performance are correct using sound processing technology. Specifically, when a beginner plays a Ryuteki, the system presents feedback on the results of the play. Then, we conducted a questionnaire survey to investigate whether this system could improve the playing technique of Ryuteki. As a result, it is expected that beginners will learn the playing technique of Ryuteki, but the improvement points of the system were also mentioned.

Humans can solve new problems by applying the knowledge of existing problems. This phenomenon is called transfer of knowledge. To investigate transfer of learned motor skills, the present study analyzed the inverted pendulum, a classic continuous motor control task in optimal control design, with a type of optimal controller. I demonstrated that, without additional motor learning, the optimal controller can be reused/applied to optimally solve other problems in the same task class, by adapting the controller function.

We have used suppressed tracking tasks and inverted tracking tasks and studied the learning processes and transfer of learning in order to investigate perceptual motor coordination. In suppressed trials, either the target or the manual cursor was suppressed for a brief period during each trial. In inverted trials, the relation between joystick movement and target movement was inverted at an unpredictable time during each trial. These tasks require learning a novel sensorimotor transformation. We have used this approach to discuss the internal models used during tracking, and their updating during motor learning. The results suggested hierarchy and modularity of the internal models.

This paper deals with estimating the knowledge about vocabulary considering contexts.

The RL framework is a well-supported framework for explaining human learning and decision-making. On the other hand, it is known that in RL, as the dimensionality of the environment increases, the learning becomes less efficient rapidly. The real-world environment has many feature dimensions, and how humans process them has been actively investigated in recent years. In this study, we created a learning task that includes dimensions assumed to be more salient and dimensions assumed to be less salient. The analysis using a hierarchical model based on the RL model showed a difference in the learning rate parameter among the dimensions. This result suggests the existence of factors that constrain the possible candidate dimensions involved. On the other hand, there were no significant differences in the learning rate parameter between conditions. Thus, it is still unclear whether or not attention was one of the factor.

We aim to distill the knowledge and to deploy the knowledge to unknown environments by transferring and combining the knowledge in multi-agent reinforcement learning. This paper proposed an implicit cooperative learning method as the way how knowledge distillation becomes available in multi-agent reinforcement learning. The proposed method makes agents learn cooperative behavior with limited information to distill the own knowledge. In addition, this paper discussed the distilled knowledge and that transfer. Concretely, under the assumption that a reward function can be divided to three terms: a term that the self agent can act to change, a term that the other agent can act to change, and a term that changes due to interactions between agents, the proposed method makes agents learn to increase the own terms of reward function and the terms for interaction without unexpected interaction. This paper investigated the performance of the proposed method by comparing with Self-other modeling and Asynchronous Advantage Actor-Critic. The experimental results showed that the proposed method uses scarce amount of information than the conventional methods and performed equals and greater, demonstrating the distillation. This paper also discussed the results of the proposed method to provide some insights and perspectives on knowledge transfer.

High-dimensional systems such as a double pendulum are difficult to control optimally due to the large number of degrees of freedom. Curriculum learning is one of the effective learning systems. It is a way to start to learn for simple task and gradually shift to a more complex task, which is an effective strategy for complex ones. In this research, we proposed a curriculum for an efficient reinforcement learning method by discovering the optimum learning initial value through a throwing task simulation.

In this paper, a query learning algorithm visualizing the prediction basis of a trained Graph Convolution Network (GCN) M whose training data is a set D of ordered trees is proposed. The proposed algorithm is based on the query learning model, one of the learning models in computational learning theory, and works with a trained GCN M as an oracle. In more detail, using a constant number of ordered trees F in D, the prediction basis of M is visualized as a representation of the ordered tree pattern by repeating queries to M as an oracle O(n²) times, where n is the maximum number of nodes of ordered trees in F. In addition, for a set D of ordered trees that match the synthetic ordered tree pattern (target pattern), we made a trained GCN M with a subset of D as training data. Then, in order to show effectiveness of the proposed algorithm, we report the ordered tree pattern (pattern for visualization) obtained by executing the proposed algorithm using M as an oracle.

Machine learning from tree structured data is studied intensively. A tag tree pattern is a rooted tree structured pattern with ordered children and structured variables. In order to represent tree structured data about complex phenomena, we propose a two-stage evolutionary learning method for acquiring characteristic multiple tag tree patterns with wildcards from positive and negative tree data, by using label information of positive examples.

Based on the query learning model, which is one of the learning models in computational learning theory, we propose a query learning algorithm visualizing the prediction basis of a trained Long Short Term Memory (LSTM) network M whose training data is a set D of sequences consisting of constant symbols (constant strings). In more detail, using a constant number of strings F in D, the prediction basis of a trained LSTM M is visualized as a linear pattern, which is a sequence consisting of constant symbols and distinct variables, by repeating queries to M as an oracle O(n) times, where n is the maximum length of strings in F. In addition, for a set D of constant strings that match the synthetic linear pattern (target pattern), we made a trained LSTM M with a subset of D as training data. Then, in order to show effectiveness of the proposed algorithm, we report the linear pattern (pattern for visualization) obtained by executing the proposed algorithm using M as an oracle.

We propose a method for discovering combinations of attributes (itemsets) against a background of statistical characteristics without obtaining frequent itemsets. The method uses evolutionary computations characterized by a network structure and a strategy to pool solutions over generations. The method directly discovers combinations of attributes such that a high correlation is observed between two continuous value variables from a database consisting of a large number of attributes as explanatory variables and two continuous value variables as objects of interest for their statistical properties. The proposed method, which seeks to achieve the discovery of small groups with statistical backgrounds from large data sets, extends the concept of frequent itemsets and provides a basis for generalizing the association rule representation.

For the estimation of brain states in spoken conversation stimuli, we conducted an experiment using three types of deep learning models (Bi-LSTM/Bi-GRU/Bi-RNN) to estimate brain activity data using speech spectrogram as speech features, and compared the estimation performance of each model. There was no significant difference in the performance of any of those models, and we confirmed that the brain regions close to the ears, which are considered to be responsible for phonological and grammatical processing, responded better. In addition, we predicted brain activity using linguistic features transcribed from auditory stimuli into text. We used RoBERTa/BERT/word2vec as a general-purpose language model to convert them into embedded vectors. In this experiment, we could confirm responses in a wide range of language areas in the brain, not limited to the peripheral regions of the ear.

Decoding inner speech from brain activity data can not just facilitate communication in patients with disabilities, but can lead to better understanding of metacognition. In a previous study, a deep learning model called EEGNet was used on inner speech decoding. However, it only achieved 30% of accuracy for a 4-class classification task. Here, the use of transfer learning is considered to be more effective. However, transfer learning has not yet been applied to inner speech. Even for EEG data in general, the effectiveness of transfer learning on various dataset has not been sufficiently verified. This study verifies the improvement of feature extraction by performing transfer learning on inner speech dataset using EEG data of different tasks and non-EEG data. The result confirms that the accuracy of inner speech is improved by transfer learning that uses data from different subjects, but not by transfer learning which uses EEG data from different tasks. On the other hand for the image dataset, the improvement of the accuracy was confirmed by freezing some layers, even though the nature of the data is different from that of EEG data.

Recognition of words in speech imagery embedded in electroencephalogram (EEG) signals is one of the challenging technologies for non-invasive brain-computer-interface (BCI). We imitate the reading process of a Rosetta Stone and develop a decoding process of a language system from brain waves (EEG). Eigen-phones, that are extracted from line spectra of EEG signals observed at Broker area, are recognized as the linguistic representation of phones. In this report we present an approach to extract not only segmental information but also suprasegmental information, or accent information.

Recently, statistical machine learning and deep learning techniques have been used to make computers learn the correspondence between imagined syllable sequences and feature sequences extracted from electroencephalogram (EEG) signals. These techniques aim to allow a computer to decode linguistic information imagined in a user’s brain. We are developing a labeling tool to efficiently construct training dataset from speech-imagery EEG. In this paper, we propose a labeling support method using the syllable similarity calculated by subspace methods and deep learning.

Speech imagery recognition from Electroencephalogram (EEG) is one of challenging technologies for non-invasive brain-computer-interface (BCI). In this report, we propose a new method to identify vowels with three features extracted from EEG signals of continuously imagined speech. The features mean line spectra detected spectral peaks using linear predictive analysis (LPA), its frequency direction derivative, and bilinear calculated tensor product of compressed line spectra. In experiments, the recognition rate was obtained approximately 73.1% by using Convolutional Neural Network (CNN). It is clear that these features are effective for vowel recognition from EEG signals of continuously imagined speech.

Brain Computer Interface (BCI) research has been started to identify recalled syllables from Electroencephalogram (EEG) during speech-imagery. Currently, it is difficult to identify the true recall duration from EEG data. Therefore, inaccurate recall data including non-recollection duration or recall sections labeled by visual determination of spectrum outline are often used to identify the recalled syllables. Because the visual syllable labeling takes a lot of time and labor, it is desirable that the process to discriminate correct speech-imagery segments has been automated. In this paper, we constructed each model consisting of speech-imagery segments and non-recollection segments to obtain the true syllable sections. We extracted the complex cepstrum from the syllable-labeled speech-imagery/non-recollection data by visual determination and identified speech-imagery/non-recollection segments using the features. Lastly, we report the classification results by 10-fold cross validation.

Although analysis of speech imagery electroencephalogram (EEG) has been actively conducted, there have been reported few numbers of results that focus on pitch accent, which is a linguistic feature of imagined speech. In this report, we propose a complex cepstrum-based accent discrimination from speech-imagery EEG signals. We first created a database containing the intervals of imagined spoken syllables that is visually labeled from the line spectral patterns of EEG signals obtained after the pooling process of electrodes. Then, we construct an accent discriminator using the complex cepstrum calculated from the amplitude spectrum from the EEG signals during speech-imagery. In the discrimination process, the eigenspaces are designed for each accent from the training data. The results of experiments using the subspace method and the tensor product-based compound similarity method showed satisfactory scores in discriminating the different types of accents of imagined two-syllable speeches.

This study realizes the Hyperdemocary Platform that is a social network that support software agents and human to collaboratively participate, discuss and make a democratic consensus. The rapid development of AI, smartphones, and the Internet, as well as the rapid changes in the environment caused by Covid-19, have made the realization of a new social system more and more realistic. Specifically, multiple agents will be stationed in social networks as the platform of democracy and act with human, mediate decisions and interactions, and support better consensus building and scalable & collective decision making. In particular, while collecting opinions and preferences, software agents are attentive to the emotions of each participant, helping humans to efficiently reach a consensus that is both proactive and satisfying. In other words, the system solves the trade-off between proactive and satisfied consensus building by humans and super-efficient and super-rational consensus building by software agents. In the SNS, social problems such as flaming, fake news, group polarization, and gerrymandering have been pointed out. Our agents will solve these problems by collaborating with human in the platform.

Deliberation on social networks is shaping the future of democratic processes and public discourse. With the increasing scale of social networks, it is however difficult to understand how deliberative processes work at scale and how they could be automatically optimised. Identifying deliberative processes in online debates and understanding their evolution could shed light on how quality-deliberation occurs and how to harness it using AI technologies. In this paper, we quantify deliberation in online debates and then propose a principled methodology to study their evolution. We start by looking at debates structured around issues, ideas, and arguments. We then analyse the time series of such utterances using natural language processing (NLP) and information theoretic methods. Finally, we conduct a social experiment to evaluate the method on an online debate. We show how the method identifies stable behaviors that reflect deliberative patterns in debates. We also show the role that non argumentative utterances have in creating feedback loops that characterise deliberative processes.

The purpose of this study is to perform link prediction, which is one of the subtasks of discussion structure extraction, with high accuracy. Link prediction is a task to predict relations of opinions. When we know the relations of opinions, we can transform discussions into tree-structure graphs and analyze the flow of discussions. In this paper, we propose a method of link prediction which uses Gated Attention Networks (GaAN). The experimental result shows that our method can predict links in discussions with higher accuracy than the existing studies.

Heated discussion or conversation in online communities interferes in smooth communications and civil settlements. To prevent such unhealthy upsurge, it is important to understand what is the feature common in the posts which are prone to trigger it. We examined whether there is a connection between heat-provoking posts and linguistic features. First, we constructed a comment dataset consisting of approximately 45,000 comments posted on Japanese Wikipedia community pages. Next, we defined "overheat" phenomenon and five features and calculated feature scores of all comments. Each comment was classified into four or two classes based on the definition of "overheat." In the analysis of comments, we compared these classes using the calculated features. The results of the analysis show that there are certain linguistic differences between these classes.

In this research, we prototyped a chatbot aimed at improving the sense of ownership of people who have not been interested in social issues so far, and by showing a familiar impact on users of social issues, it will lead to an increase in the sense of ownership. A hypothesis was made and a proof experiment was conducted. In the empirical experiment, the results suggesting the effectiveness of the proposed method that supports the understanding of the problem background of social problems and shows the familiar impact on users in order to get them interested in social problems.

In public discourse, citizens are expected to discuss referring to diverse common goods such as utilitarianism, equality, and maximin principle. The discourse quality should be evaluated from multiple viewpoints. The present study conducted group discussion experiments on the treatment of low removed concentration soil outside Fukushima. We evaluated the discussions regarding three types of common goods with the revised Discourse Quality Index (DQI) and scores of the participants and observers. There were two conditions: one with information about Fukushima and the other without the information. The study examined how this different information treatment would affect the evaluations on common goods. Results from the renewed DQI indicated that the number of arguments about the maximin principle was less than on the other common goods in both conditions. Results from scores of the participants and observers found a significant difference in the evaluations of the maximin principle between conditions. Moreover, a significant difference was found in the non-informed condition, indicating that the maximin principle was less discussed than utilitarianism, while it was not found in the informed condition. This study provided the empirical findings to help make AI enable to evaluate democracy.

In most discussions to make consensus formation, the discussion topic includes multiple issues, and the result of discussions are based on these issues. We conducted discussion session in International Convention and Tourism Symposium and the discussion was done on invitation and holding of international conventions with more than 30 convention staffs and local government staffs in Japan. In the symposium, a social experiment was conducted to see how discussions using both online-verbal discussion and online discussion/consensus formation support system D-Agree to solve convention business problems can bring about changes in participants' thoughts. This paper shows the comparison between the result of pre-survey and post-survey in terms of changing perceptions regarding motivations for participation when in the position of a buyer and thoughts about the convention business as a seller.

In this preliminary work, we present an idea of abstract framework of cooperation for handling argumentations by groups. We formulate group formability semantics in this abstract framework, paving a path to further study.

Dissolved gas analysis (DGA) is widely used as a method to diagnose internal abnormalities in electrical transformers, such as overheating and partial discharge. While electric power companies conduct inspections and repairs according to the diagnosis results based on DGA, more efficient maintenance work based on fault sign detection is required in terms of stable power supply and cost reduction. This paper shows the results of a basic study on the prediction of fault signs in oil-filled electrical transformers using DGA. In order to predict the fault signs in oil-filled transformers, the distance to the decision boundary and the classification probability generated by multiple machine learning methods were analyzed.

This paper proposes a new fault diagnosis method that combines Multivariate statistical process control (MSPC) and a linear non-gaussian acyclic model (LiNGAM), referred to as MSPC-LiNGAM. MSPC is a widely adopted process monitoring method based on principal component analysis (PCA). In MSPC, T² and Q statistics are used as monitoring indexes for fault detection. Contribution plots based on T² and Q statistics have been proposed for fault diagnosis. However, contribution plots do not always appropriately diagnose causes of faults. In this study, a new fault diagnosis method based on MSPC and a Linear Non-Gaussian Acyclic Model (LiNGAM) is proposed. In the proposed method, referred to as MSPC-LiNGAM, the causality among the T² or Q statistic in addition to process variables is calculated by LiNGAM without prior knowledge of processes, and process variables that have the strength of causality to the T² or Q statistic are identified as candidates of the causes of the fault. The proposed MSPC-LiNGAM was applied to a simulation data of the Tennessee Eastman (TE) process. The result showed that the proposed method appropriately diagnosed faults even when the conventional contribution plots did not correctly identify causes of faults.

Nerve cells are composed of cell bodies and neurites, and it is known that most of the motor functions of cells are produced by the expansion and contraction of neurites. Therefore, in order to estimate the degree of activity and motor function of the cells themselves, it is necessary to perform a detailed analysis focusing on the neurites. However, it is difficult to define the neurite region strictly, and in many cases, it is judged by the human eye, so there is a problem that only subjective and non-quantitative evaluation can be performed. Therefore, in this study, we attempted to realize a more quantitative and objective evaluation method focusing on the protrusion region by mechanically detecting the protrusion region by utilizing the rule-based approach. Specifically, polar coordinate transformation was applied to the masked cell image, and the protrusion region was automatically detected by defining the protrusion region from the unevenness of the polar coordinate graph. As a result of simulation experiments on some annotated mask data, it was confirmed that appropriate protrusion detection and quantitative evaluation of protrusion movement based on the detection results were performed.

In this paper, we propose an unsupervised anomaly detection method using user feature embedding by graph convolution to detect insider threats in the field of cyber security.In recent years, research on insider threat detection using machine learning has been conducted in the field of cyber security.In general, supervised learning is used for detection. However, in real-world data, only a few of them have correct labels. Therefore, supervised learning is difficult.In this study, we used unsupervised learning for insider threat detection.And we construct a graph from the dataset and show that the accuracy can be improved by embedding features using graph convolution.For evaluation experiments, by analyzing the dataset, we discovered differences from real-world data and defined a more realistic problem setting.

Precisely characterizing cell activity is an important factor when evaluating potential cures for life-threatening diseases such as Alzheimer’s and cancer. It requires precisely registering the time-dependent cell location and especially cell morphology. In bright field black and white images, many cultured cells are especially challenging to distinguish from the background, contaminates, and within adhesive clusters of cells. A mixed approach, consisting of deep learning and physics-based methods, is used to estimate the cell response within a variety of in-house produced datasets. In this research, the morphology and motility evolution of human neuroblastoma, SH-SY5Y cells, in low-contrast time-lapse observations was evaluated. Temporal cell nuclei analysis was performed to aid cell separation from clusters and contaminants. This improves the segmentation results to achieve high mAP (>0.95) at high values for IoU (>0.8) and permits comparable cell activity characterization.

We developed a training interface that visualizes the virtual agent's subjective values and emotional appraisal process in a multi-issue ultimatum game in order to improve the ability of mental states inference and win-win calculation. Experimental results showed that the training interface contributed to improve people's ability of win-win calculation.