For wider use of deep learning (DL) in society, deeper understanding of fundamental principles underlying various DL architectures is needed so that the users have better control over what they are actually doing with DL technologies. The deeper understanding is also likely to be useful for developing more environmentally friendly learning methodologies. As a preliminary step toward this goal, we study fundamental properties of signal propagations in artificial deep neural networks in this paper. More specifically, we show that there is a strong analogy between the signal propagation process in appropriately initialized fully-connected/convolutional deep neural networks and the dynamics of the so-called "absorbing phase transitions (APTs)" which can be found in some physical systems driven far away from equilibrium. We discuss, with numerical results on the signal propagation process, how these neural networks can be placed in a context of the theory of APTs and what theoretical/practical implication can be gained beyond the well-known mean-field theory.

The recent high performance of many deep learning models heavily relies on the massive amount of labeled data for training. Still, the correct annotation may be expensive or even impossible in practice. Unsupervised domain adaptation aims to train a robust classifier when we have access to unlabeled samples from a target domain and labeled samples from a source domain, which has been intensively studied in the literature. However, it is often relatively easy to obtain additional noisy labels from the target domain by, e.g., heuristic labeling, crowdsourcing, and the prediction from a preliminary model. This paper considers a novel problem setting where the target samples are equipped with noisy labels and proposes a method to incorporate the noisy target labels in the generalized label shift framework. We evaluate its performance via thorough experiments on several benchmark datasets and show that it can help transfer knowledge from the source to the target domains.

Recently, there has been growing interest in Federated Learning (FL) for learning from distributed data. While FL has advantages such as privacy protection and traffic reduction, it is difficult for existing FL models to preciously characterize the data distribution of each client. In this study, we consider an autonomous federated learning scheme that allows to maintain both client and general models to capture the individual and common data for each client, respectively. This scheme enables to extract features of client's own data. Preliminary experiments on the image benchmark data demonstrate the usefulness of the proposed method in terms of the performance and data feature extraction.

Learning is the result of thinking experience. Thinking experience, then, depends on the characteristics of the domain of learning. Therefore, the domain model plays importance role in the design of technology enhanced learning. In this paper, as an example of domain driven design of technology enhanced learning, open information structure approach is introduced. As a concrete example technology enhanced learning, problem-posing exercise of arithmetic word problems is explained.

Our goal is to build a Peer companion robot for learning support. For this purpose, it is essential to understand the knowledge of Peer-ness induction by robot actions. Therefore, we experimentally evaluated the validity of long-term interaction by presenting robot actions that follow the psychological model of the "Peer-ness" mechanism obtained in our previous study to induce a "sense of competition," and whether this could induce "Peer-ness”.

This paper describes a model design approach for Web-based investigative learning, in which learners are induced to decompose a question into sub-questions to be further investigated. Such question decomposition plays a crucial role in creating a learning scenario implying the questions to be investigated and their sequence. It also contributes to obtaining deeper and wider knowledge about the question. In this paper, we discuss how learners should conduct the question decomposition in a reasonable, sufficient, and diverse way, and how to support it.

A difficulty of writing tests is that given scores depend on rater characteristics, such as severity and consistency. To resolve this problem, the generalized many-facet Rasch model (GMFRM), an item response theory model that considers such rater characteristics, has been proposed. When applying such an IRT model to datasets comprising results of multiple writing tests administered to different examinees, test linking is needed to unify the scale for model parameters estimated from individual test results. In test linking, test administrators generally need to design multiple tests such that examinees or raters partially overlap. However, preparing common examinees and raters is often difficult in actual testing environments. Therefore, in this study, we propose a novel method to link the results of the writing tests, which are obtained by applying GMFRM, using recent neural automatic scoring technology. Experimental results show that our method realizes the test linking without common examinees and raters.

In learning, it is important to reflect on one's own answers to deepen understanding and to perform trial-and-error independently. We have developed a learning support system for vectors in mathematics that visualizes errors by converting mathematical expressions input by the learner into figures that can be operated according to the constraints of the mathematical expressions, thereby providing an environment in which the learner can perform trial-and-error operations. However, previous learning support systems stopped at presenting error visualization to the learner, and the operation of figures and the confirmation of feedback were dependent on the learner. Therefore, there was a possibility that some learners would not check the feedback given to them and would not look back to see what errors they made in their own answers. In this study, we propose a learning support that requires the system to perform operations by presenting a triangle as a model of the relationship between vectors in the synthesis of vectors in the problem, in addition to the conventional learning support. Since the learner is required to perform the activity of forming a triangle based on the generated vectors, the learner can reflect on the errors in his or her own answers, and can perform trial-and-error operations. This paper describes the implementation of the proposed learning support in the system.

Techniques for detecting Adversarial Examples applied to the input data to a neural network include metric-based approaches, denoting approaches, prediction inconsistency-based approaches, network invariant checking approaches, etc. The most important of these approaches are the metric-based approaches. This presentation describes an implementation method for practical use of the network invariant checking approach (NIC method), which has been reported to have the highest detection rate among these approaches, and reports that it was actually able to detect adversarial perturbations with a high detection rate. The results of a discussion on the reasons for the high detection rate of the NIC method are also reported.

AI software differs from traditional software in that it is generated inductively from training data. Therefore, it is essential to prepare high-quality training data. We have previously proposed a method to evaluate the quality of data using a variational autoencoder. It is a laborious and challenging task for users to manually map latent variables to attributes, which consumes a lot of time and makes it difficult to quantify the attribute values. In this study, we propose a semi-supervised representation learning approach that can automatically and quantitatively predict attribute values from latent variables by enhancing the variational autoencoder. Our proposed method incorporates a term in the loss function that includes the coefficient of determination, which measures the goodness of fit of the regression equation. The purpose of this is to predict the attribute values from latent variables through regression analysis in the labeled data, which is a portion of the dataset. The method is then trained to increase the coefficient of determination. The results of applying the proposed method to a dataset of handwritten characters on forms indicated that it is an effective method for objectively evaluating the quality of the dataset.

Recently, many visualization methods have been presented that target information on the model itself. On the other hand, there are few methods to visualize information about the workers (annotators, model designers, and end-users) of the model. The active intervention of workers in the modeling process is effective in improving the accuracy of models, and visualization of workers is considered useful for understanding the properties of models in detail, evaluating adjustment work, and suggesting effective improvement measures. Therefore, we propose a visualization tool that focuses on the model modification history and the objectives of individual tuning operations as information about the operators. Our tool calculates the differences in data, model structure, and optimization algorithms during model tuning, and visualizes them together with the intent of the change. We present the visualization results of the history of training and testing a model for image classification.

Currently, evaluation indicators, such as accuracy, precision, and recall, for datasets are widely used for evaluating machine-learned models represented by deep neural networks, but it is difficult to guarantee performance for unseen data not included in the datasets by such evaluation indicators. In this presentation, we explain how to use the noise-added generalization error bounds as an evaluation indicator to probabilistically guarantee performance (incorrect-rate) even for such unseen data, based on statistical learning theory, and report experimental results for demonstrating the effectiveness of the indicator. Here, the generalization error is the expected value of the incorrect-rate of the output of a machine-learned model for all input data selected according to a probability distribution. In general, it is difficult to exactly compute the generalization error because of the innumerable of input data, but there have been a lot of related works on bounds of the generalization error. We apply well-known theorems on training-set-based generalization error bounds called PAC-Bayesian bounds, as testing-set-based bounds, to compute bounds close to generalization errors.

In the operation of AI system using machine learning technologies, the quality and performance of the system may deteriorate due to changes of data in operation. In particular, the change in data distribution, called concept drift, is one of the main causes of performance degradation. In addition, in the operation of upcoming AI system, there would be the cases where the training data used before the operation cannot be reused because of their privacy and security concerns. In this paper, we arrange concept drift detection and adaptation technologies and unsupervised domain adaptation technologies that we have introduced in JSAI2020, JSAI2021, and JSAI2022 by showing their effectiveness to maintain quality and performance of the AI systems in operation. Furthermore, we also introduce the recent research on “Test-time adaptation techniques” which adapt machine learning models online without reusing training data, effectively addressing new problems that may arise during AI operation.

The aim of this paper is to consider the relation between affective understanding and intellectual understanding, drawing on Kiyoshi Oka’s work. Based on his own experience of doing mathematics, Oka emphasizes the role of affectivity in understanding, which guides one to grasp the meaning of the object as a whole beyond/before the detailed intellectual understanding of it. Viewed from the current emotion research, Oka’s claim seems to correspond to the bodily perception theory of emotion. In this theory, emotion is regarded as a cognitive capacity to apprehend the value aspect of the given situation, as perception is a cognitive capacity to capture the fact aspect of the situation. In my view, the capacity of articulating an experience in words seems to integrate emotion and perception into the whole process of understanding. Affective understanding and intellectual understanding are mediated by our capacity of verbalizing what is experienced as an emotion.

“Joh-cho” is the harmonious and integral combination of intellect, emotion, will, and sensation, belonging to neither “thing” nor “mind”.We must acquire “Joh-kai”, an affective way of understanding, in order to understand “Joh-cho”.Authors believe that embodiment supports “Jo-kai”, considering phenomenological concepts such as “Hyoh-joh”, “Kinesthetic Melody”, and so on.On the basis for that concepts, we are able to acquire “Jo-kai” by seeing through our own body, not only with our own eyes.Therefore, it is necessary for acquiring “Joh-cho” and “Jo-kai” not only to study through writing about that concepts, but also to encourage our own body to be a medium of seeing, which authors call that “emergence of Mi-ugoki”In this paper, we also introduce a piece of the practice with an application which encourages emergence of “Mi-ugoki” we are working on.

While engaged in the digitization of cultural properties, a particular part of the image of the requested material sometimes becomes a matter of concern during the image confirmation process. This interest continues for a long period of time. This may be similar to what Kiyoshi Oka means when he says, "Rather than searching for something, I continue to be interested in it. Interest also seems to lead to the discovery of latent knowledge. This paper attempts to reflect on the author's work experience and analyze it.

n recent years, cognitive science has focused on the embodied nature of consciousness and its interaction with the environment. Francisco Varela referred to this philosophy of cognition as "embodied mind," and emphasized that cognition must be understood in relation to the individual history of the body and its organic environment, as well as the situation in which the organism is acting. If we interpret the embodied mind more broadly, cognition should be seen within the comprehensive environment of the cognitive subject's bodily habits, language use, and ecological environment. In other words, consciousness is a phenomenon that could be understood from the perspective of "co-evolution with civilization." This paper attempts to identify how the mind functions in Japan from Soseki Natsume's works and theories. In Soseki's works, consciousness is a combination of information processing and bodily sensations, which suggests a different way of understanding consciousness from the current cognitive science.

The objective of this discussion is to depict a preferable and normative science of design to understand the wisdom employed in activities where artificial things including intelligence are created. The primitive nature of designing is to device self-consciously a course of action to live well. The cognitive processes in designing are hypothetically depicted from the substantial, somatic, and conceptual aspects. With this picture, a method of understanding wisdom in designing is discussed in association with the world view of a mathematician OKA Kiyoshi. Oka claims that an individual human is living not in the material world but in the world of the individual mind and emphasizes the importance of understanding the things as living matters as well as substantial matters. This discussion is also focusing on the concept of intellect, emotion, and volition, and the concept of levels of creation of the nature, i.e., living, life, and self-consciousness of living.

The future is just around the corner when self-driving cars, robots, AI, and other machine intelligence will surround us. What kind of environment will we be in when that intelligence and capability, however simple and limited, crosses a certain amount and becomes part of the landscape? As personal computers and the Internet spread worldwide, what kind of intelligence will the new tools and environments bring out in humanity? Many of these questions were discussed. How will machine intelligence surround the environment, and what kind of intelligence will be drawn from it? To answer this question, we are conducting a project called "Smart Cities, Fungi, and Buddha." We are interviewing people in various fields, referring to fungi as a symbol of intelligence different from that of humans and Buddha's intelligence as intelligence that transcends humanity. By referring "Jyoucho" described in Kiyoshi Oka's "A Life of Mathematics" and animism in haiku and Alaya-vijnana as a way of creativity, this paper envisions a new way of depicting the world that deviates from the modern city that symbolizes conscious and rational thinking.

Intelligence does not consist of only information, but also activities in the body and actions from the world dynamics. Intelligence is dynamical and hierarchical system. In the process, recognition is formed, and becomes connected with five senses, and finally vanished. Self-image and world-image are formed concurrently, and relation between them are structured. The relations are triggers of intelligence. In the session, it is described how an integrated flow in the hollow structure of agent architecture, which includes information flow, action flow, and inner flow in the body, produces recognition, emotion, and will.

The notion that researches and academic disciplines should be done by scientific methodology has become stronger in recent years. However, its precise doctrine concerning objectivity and universality tends to deprive researchers of attitudes of directly dealing with actuality in human being’s living. That might downgrade the quality of studies on human intelligence. Kiyoshi Oka, one of the mathematicians representing Japan, has brought the concepts of living, life, and wisdom, and thereby presented a warning towards academic activities. The purpose of this paper is to discuss ideal attitudes of doing research on intelligence and propose a new style of research, studies on/with first-person’s view, as a method of coping with tacit-ness, situatedness and symbol grounding which are all significant issues, but difficult ones to bring to realization in AI systems

This study demonstrates whether the financial text is useful for using stocks in the tactical asset allocation method. This can be achieved using natural language processing to create polarity indexes in financial news. We perform clustering of the created polarity indexes using the change point detection algorithm. In addition, we construct a stock portfolio and rebalanced it at each change point using an optimization algorithm. Consequently, the proposed asset allocation method outperforms the comparative approach. This result suggests that the polarity index is useful for constructing the equity asset allocation method.

In recent years, with the development of machine learning, there has been an increase in research on the application of deep reinforcement learning to optimization problems. In this study, we applied reinforcement learning (Actor-Critic) and Pointer network to solving the Knapsack problem used in cryptography, and applied reinforcement learning (Actor-Critic) and Pointer network to solving the TSP. We propose a method for solving the trapdoor (Knapsack problem) of the Knapsack cipher based on the work of Google Brain, which applied reinforcement learning (Actor-Critic) and the Pointer network to solve the TSP. Specifically, we solve random, hyper-accretive, and modulo hyper-accretive trapdoors, obtain exact solutions, and compare the results with the LLL algorithm. The results showed that both problems can be solved up to 30 dimensions, and the LLL algorithm was more accurate than the LLL algorithm for some problems, but the LLL algorithm was basically better for higher dimensions and lower densities.

Patent literature consists of three parts: claims, descriptions, and drawings. The examiner must retrieve the database using classification symbols and text to check whether claims include existing documents. This paper aims to treat a patent drawing with variations as structured graphs and to establish an efficient method for similar floor plan retrieval in patent examination. First, we labeled the patent drawings with room types. Second, we stipulated the connection relationships of the rooms based on rules and converted the labeled patent drawings into graphs. Third, by reflecting claims in the graph similarity based on the maximum common subgraph, we implemented a similarity floorplan search that considered the unique circumstances of patent examination. Finally, using an existing reviewed literature, we verified the usefulness of the proposed method in terms of similarity-based patent sorting results and computational speed.

In the retail industry, there is a growing need for AI to identify products on product shelves in order to improve the efficiency of product management and planogram. Image identification technology, which identifies individual product images detected by object detection from product shelves by matching them with database images, is required to distinguish hundreds to thousands of products, including those with similar designs, under different lighting environments and display conditions for each store. In addition, it is required to be able to follow frequent product design changes. Furthermore, the number of images that can be used for learning is usually limited for each product. Deep metric learning (DML) is an effective approach to such problems. In this study, we propose data augmentation from one-shot images, clusterwise attractive/repulsive loss, epoch-by-epoch pairwise semi-hard negative mining, utilization of self-attention mechanism in backbone CNN, etc., and tried to acquire performance. Recall@5 performance achieved about 97% for the trained category and about 95% for the untrained category. Due to the characteristics of DML, it is possible to improve the efficiency of correcting mismatching.

Penetration testing assesses the vulnerability of a target system by exhaustively attempting known attacks. Efficient penetration testers carefully investigate the target system to minimize attack attempts. In this study, we propose a penetration tester based on a neural agent that efficiently finds the optimal penetration strategy using deep reinforcement learning based on partially observed Markov decision processes (POMDPs). In addition, while the baseline neural agents are based on GRUs, we propose a system with a variant of Transformer, called GTrXL, which is expected to solve the state prediction problem in neural agents from partially observed text using NLP techniques. We have conducted several experiments against real linux-based systems through a wide-known autonomous penetration testing tool, called DeepExploit, as a part of the environment. We have succeeded in demonstrating the superiority of our proposed GTrXL-based agents against cutting-edge previous studies

In this study, we examine what strategies survive under infinite group dynamics when players cause implementation errors in an iterated prisoner's dilemma. In the previous studies, the Win-Stay, Lose-Shift (WSLS) strategy, which is a representative strategy can be the most abundant in the presence of implementation errors among memory-one strategies. In contrast, we use replicator-mutator dynamics to examine the consequences in a strategy space represented by finite state automata within two states. As a result, we observe a Unilateral Defection, Mutual Defection (UDMD) strategy that behaves effectively when players cause implementation errors, and clarified the survival process.

In this study, we propose a multiplicative weight update algorithm that utilizes mutations in two-player zero-sum extensive-form games. These games are important models for decision-making under imperfect information. While equilibria in these games can be computed using linear programming, it becomes challenging to handle large-scale games such as poker. To address this issue, learning algorithms for finding an (approximate) equilibrium have been proposed. However, most of the existing algorithms converge to Nash equilibrium through time-averaged strategies. In normal-form games, it has been shown that introducing mutations allows for learning equilibrium strategies without taking time averages. Inspired by that, we propose the Dilated Mutant Multiplicative Weight Update with the introduction of mutations in extensive-form games. The experimental results show that the proposed method can learn equilibrium strategies without computing time averages for multiple settings.

This work addresses learning online fair division, wherein the values of items that arrive sequentially are not directly observable, but instead the noisy, estimated values are observable when we assign the items. We consider the problem as computing market equilibria in linear Fisher markets where agents have additive utilities. In such markets, the static allocation simultaneously achieves envy-freeness and Pareto optimality by maximizing Nash social welfare, assuming that items are divisible or can be allocated randomly. However, this fact is no longer valid in online settings. To this end, we have developed online algorithms that combine dual averaging with multi-armed bandit indices. Through dual averaging, our algorithms gradually learn the values of arriving items via bandit feedback. As a result, the algorithms asymptotically achieve the optimal Nash social welfare. We also empirically demonstrate the superior performance of the proposed algorithms in synthetic and empirical datasets.

Widely used as a tool for comparing probability distributions, the optimal transport (OT) theory is very important in many machine learning tasks. Sinkhorn's algorithm successfully reduces its compuational cost from a cubic complexity to a quadratic one. Nevertheless, popular approaches of distribution comparison with OT on feature sets of different sizes could not support GPU parallelization. In order to overcome this difficulty, we propose the basis optimal transport which provides a translated OT problem with distributions of fixed sizes. Futhermore, we propose a deep dictionary learning framework for translating a given OT problem into our proposed basis optimal transport problem to make it solvable with GPU-based Sinkhorn's algorithm. A great reduction of computational time cost is reported according to our numerical experiments for computing the Wasserstein distance on datasets with size-variable distributions.

Multiagent pickup and delivery problem has been studied for autonomous carrier robots in warehouses and autonomous car operation. A fundamental solution method resolves the confliction among endpoints of tasks' paths and avoids the deadlock of paths. While several heuristic techniques have been proposed to improve the solution method, there are opportunities to investigate the integration of those techniques. As a study toward the integrated efficient solution methods, we integrate two efficient techniques that reduce the redundancy in warehouses' space utilization and tasks' paths. We experimentally present the effect of the proposed approach and consider the possibility of generally integrated efficient methods.

This study is a proposal about the design patterns of character AI in 3D action games. Currently, there is no fixed method for the development of character AI in digital games. Therefore, this study proposes a method using pattern language as a development method for character AI. This method consists of four processes: “sampling character behavior from games “, “creating pattern language”, and “implementing pattern language”, “verifying pattern language”. A list of character AI design patterns that character AI should follow was created based on character behaviors that users felt unnatural. Next, the behavioral patterns of compliance and non-compliance with each norm of the list were implemented on a game engine (Unreal Engine 5) and compared each other. The implementation of four design patterns in the list was explained and their effectiveness showed in this article.

Early turnover rate has remained around 30% for the past 30 years. The main reason is that personal relationships in companies were not good. To the best of our knowledge, there is no previous study that considered early turnover from the viewpoints of both formal and informal relationships. The aim of this paper is to propose a company organizational structure that would prevent early turnover considering both of these relationships. We built a simulation model to infer early leavers from formal and informal relationships. Based on this model, we propose a company organizational structure in which one team is formed by younger members of the existing workforce, and new employees are assigned to the team. The simulation results showed that the proposed structure was more effective in reducing early turnover than the existing top-down corporate structure.

Indirect reciprocity is a mechanism for the emergence of cooperation in society. This mechanism enables us to explain cooperation provided by a stranger, and is considered to be established by transmitting cooperative actions in the past to a third party as a good reputation. However, in real society, there may be other factors to motivate cooperation. In this study, we consider income as an index of wealth and analyze the impact of income information on the emergence of cooperation. For this purpose, we provide a model where agents play a repeated prisoner's dilemma based on income information. Then we evaluated the ratio of cooperative behavior chosen in the game through simulations with the model. As a result of the simulations, we cannot observe the emergence of cooperation in the setting when income information is given.

In videoconferencing, it is difficult to establish the methods by which people's opinions change when an agent intervenes. In particular, it is an important research issue to understand which intervention methods brought about changes in people's opinions. In this study, we investigate the effects of an intervention on agents' discussions in videoconferencing. To this end, we first implement an agent that expresses opinions opposite to those of others and then test its effectiveness. The experiment was conducted with 14 subjects divided into two groups: one without an agent (hereafter referred to as the "no agent intervention" group) and the other with an agent (hereafter referred to as the "with agent intervention" group). Although the agent's intervention did result in a change in opinion, the results of testing the hypothesis based on statistics did not reveal the effect of the intervention. However, the small population size of this survey needs to be more appropriate for testing hypotheses based on statistics. Therefore, it is necessary to conduct further experiments with many participants as a future issue.

To decide the next move in Shogi, it is necessary to narrow down the candidate moves from several possible moves, read through the moves, compare the read-throughs, and make the move that is judged to be the better move. In general, there are many possible moves in Shogi, and a Shogi beginner does not know which move to consider. Therefore, existing Shogi AI assists the user by presenting candidate moves. The suggestion of a candidate move is a specific move such as "move the 28 rook to the 68 square". However, the assistance with candidate move suggestions may encourage the user to abandon thinking by simply playing the suggested move instead of reading the future development. Such excessive assistance may cause the learner to lose the ability to self-learn, which is called the "assistance dilemma". In this study, we propose a system in which only pieces that should be moved, such as "move the 28 rook," are presented, and where to move each piece is not presented. Through experiments, we have compared the proposed system with existing assistance systems and investigated its potential to promote thinking and solve the assistance dilemma.

The objective of this study was to build a more efficient method of text input in an MR environment. To achieve this goal, we tested a system that combines four functions or forms of input: hand tracking, gesture input, 3D keyboard, and direct touch to the keyboard. The user evaluation results showed that the SUS rating was generally good, but there were some problems such as mistaken input and high input difficulty.

Cooking by ourselves is important to manage our health. But it’s difficult especially for beginners to reproduce recipes. As a solution to this issue, systems supporting persons who are cooking by feedback from their activities are considerable. In this study, we propose a method of real-time cooking activity recognition using accelerometers to realize such systems. Our proposed method classifies the subject’s activities during cooking of curry to 7 classes by convolutional LSTM with frequency of one time per a second. We confirmed a performance of our proposed method was 86.39% in batch inference, 61.32% in real-time inference as results of evaluation with macro-average of recalls.

In cooking, it is important to both focus on each cooking process and to consider the entire cooking process. However, since many cooking recipes are expressed using text and images simply, it is difficult for a beginner to grasp the entire procedure from a bird's-eye view. In this paper, we describe an interactive visualization system that converts cooking recipes into flow charts. Our system, named GraphRecipe, is realized as a single page application and consists of three parts: cooking images, flow chart view, and cooking information. The flow chart expresses the entire procedure in a concise format, and the progress of the current cooking process and the details of the next work can be grasped immediately.

We investigate the expression method for a pet-type robot to persuade a person. There is a study in which a robot tries to perform some actions to persuade a person. We created a persuasive expression that combines several actions and emotions. The actions were classified into three meanings: "calling," "guiding," and "pointing" for the pet-type robot. For emotions, six emotion patterns were created by combining various parts of the robot such as the face and tail, voice, and speed of movement. By conducting an experiment in which 12 patterns combining movements and emotions were compared by displaying the robot using MR in a real space, we evaluated the expression methods used to persuade. The experimental results showed that emotional expressions such as "sad" and "troubled" were effective when the pet-like robot was used to encourage the subjects to throw away the trash. On the other hand, emotional expressions that gave the impression of "happy" or "angry" were not effective. In addition, there was little difference in evaluation based on the combination of behaviors.

When developing commercial games, difficulty levels are carefully designed to satisfy many game players. Inappropriate difficulty levels can cause stress to players, leading them to quit the game. In recent years, a system called "dynamic difficulty adjustment," in which a computer analyzes the player's playing situation and adjusts the difficulty to an appropriate level, has been attracting attention as a solution to this problem. In this study, we use biometric sensors to measure and analyze players' biometric information during game play to clarify the relationship between game development and emotions. Furthermore, we will develop a game that can estimate changes in emotion in real time using biometric sensors, automatically adjust the difficulty level of the game according to the emotion and verify whether the adjustment is appropriate for the player. The results obtained from the experiment did not provide definitive indicators or evidence of usefulness, but suggested the possibility that the emotions that are likely to be expressed differ depending on the degree of difficulty and personality. The results obtained from the experiment revealed that the emotions that were likely to be expressed differed depending on the game scene and the difficulty level.

We investigate communication using avatars projected on a spherical display between a user using a head-mounted display from a remote location and a user near a spherical display at a local location. The remote participant becomes an avatar on the local spherical display and communicates with the local participant by pretending to be a different person on the spherical display. The remote participant can look around the site using an all-sky camera that can capture a 360°view of the surrounding space. The results of the experiment with the subjects showed that it is possible to convey emotions to others through changes in facial color and facial expression, and that additional efforts are needed to communicate by pretending to be a different person.

Emotional contagion is the exchange and transfer of emotions to others. Dyadic dynamics in a group setting received little attention, especially when combined with emotional contagion as the contagion of emotions is very low due to the low number of group members (i.e., 2). In this paper, we perform a preliminary experiment by using EDA and ECG sensors to monitor 2 subjects while they are interacting with each other. We choose the subjects to be a married couple in order to ensure strong emotional contagion, thus, interesting insights can be obtained. Subjects answer the Emotional Contagion Scale (ECS) questionnaire, Facial Expressiveness Scale (FES) questionnaire, and General Health Questionnaire (GHQ-12). Our results show that the subject with a high score of FES is easily and mostly aroused, thus leading the emotions of the other subject. The emotional contagion phenomenon is clearly shown in the ECG measurements too. ECG of the subject with high FES leads/precedes the other subject, thus changes (in the form of replications) can be seen almost instantly. However, changes take longer when its related to EDA measurements. Emotions transmit from the charismatic personality to the empathetic personality, thus empathetic personality's ECG and EDA sensory information are imitating the charismatic personality's sensory information after a certain delay.

The advent of deep learning models has made considerable strides in tasks related to navigation in the real world such as object detection and path planning. It has also led to the development of a more complicated task of visual-linguistic navigation (VLN) i.e., dialogue for navigation. Among VLN variations, outdoor scenes are significantly more difficult than indoor because of the randomness inherent in an uncontrolled environment. Outdoor VLN is also said to be closer to the reasoning required in the real world. However, the datasets available for Outdoor VLN tasks have been focused mainly on judging spatial reasoning abilities. This is not enough to create systems that work in real life as there is a need for commonsense reasoning abilities i.e. social and event-based reasoning. We create a small benchmark commonsense reasoning-based dataset and evaluate the performance of state-of-the-art VLN models on it. From our findings, we show that there is a need for commonsense reasoning-based datasets.

In recent years, Graph Neural Networks (GNNs) have demonstrated significant advances in accuracy for various graph-related tasks. However, GNNs still fail to achieve high performance in graph classification tasks. One of the primary reasons for this is that GNNs cannot learn key subgraphs that contribute to the prediction. Some research on identifying key subgraphs has been conducted within the field of Explainable AI (XAI) in graphs. Especially explanation confidence (EC) is an important evaluation method for XAI models of GNNs. In this paper, we propose a novel method for learning GNNs that incorporates Explanation Confidence (EC). We demonstrate that the proposed method performs as well as or better than conventional methods in graph classification experiments.

(1) The objective of this study is to generate artificial intelligence (AI)-based models to predict the activation state of the molecular pathway networks. (2) Since the activity of the epithelial-mesenchymal transition (EMT) is involved in anti-cancer drug resistance and cancer stem cells, we used AI modeling to identify the cancer-related activity of the EMT-related pathway in datasets of gene expression. Molecular network pathway analyses were performed on the gene expression data of diffuse- and intestinal-type gastric cancer. A dataset of 50 activated and 50 inactivated pathway images of EMT regulation by growth factors pathway was modeled by the DataRobot Automated Machine Learning platform. The AI application created a Light Gradient Boosted Trees Regressor model to predict the activation state of the EMT pathway. The model was validated with 10 additional activated and 10 additional inactivated pathway images. Our approach holds promise for modeling and simulating cellular phenotype transition.

While the world of bridging anaphora resolution (BAR) is ruled by the multi-task algorithms, they are used to only perform BAR and not coreference resolution (CR). Furthermore, their pipelines are very demanding in terms of computational and spatial resources. We therefore propose to use the work of Dobrovolskii (2021) on CR and Yu and Poesio (2020) on multi-task BAR to create a multi-task and efficient algorithm that will be able to resolve both BAR and CR at the same time. We also demonstrate that, on top of its efficiency, the results of our algorithm are still competitive both in BAR and in CR with recent works.

Machine learning (ML) techniques such as neural networks (NN) and Gaussian process regressions (GPR) are now widely used in diverse applications. While each technique has pros and cons, they are all challenged when faced with high dimensionality of the feature space or low and uneven data density. We will demonstrate how combining them with high-dimensional model representations (HDMR) results in methods better apt to deal with these issues. HDMR-NN, HDMR-GPR combinations and NN with HDMR-GPR neuron activation functions will be presented with examples ranging from computational chemistry to quantitative finance.

Visual explanation methods, such as CAM and Grad-CAM, have been proposed to visualize and interpret the decision-making of CNNs. Recently, there are some other works that not only aim to provide better visual explanations, but also to improve the performance of CNNs by using visual explanations. In this work, we propose a network architecture — MANet that generates visual explanation during the inference process using a mixed attention module for adaptive feature refinement and also uses the generated attention map to improve network performance on image recognition tasks. Experimental results show that our proposed MANet achieves better visual performance and outperforms the baseline models on both image classification and object detection tasks.

Multi-modal personality traits recognition methods recognize human personality traits to improve the quality of human-computer interaction, which has attracted increasing attention in recent years. However, current methods fail to remove noise and cannot align different modality features in the feature fusion process. To solve the above problems, we propose an emotion guided multi-modal fusion framework for personality traits recognition. Inspired by the close relationship between emotion and personality, we design a novel emotion-guided multi-modal fusion mechanism, which is expected to enhance emotion-related features by emotion-level alignment and pay less attention to irrelevant features to remove noise. Extensive experiments show the effectiveness and robustness of our model.