Recently, ICT education has been introduced to the educational field, and the research and development of educational support robots have attracted attention. In this study, we focus on a perplexion estimation method for educational support robots that estimates perplexion based on learners’ facial expressions. The accuracy of a conventional method for estimating perplexion from facial expressions at time t, is not yet practical, and we believe that the accuracy of the method can be improved by using not only facial expressions at time t but also time-series data on changes in facial expressions between t-1 and t. This study focus on a perplexion estimation method for educational support robots that estimates perplexion based on facial expressions of learners. In this paper, we verify the accuracy of the spatio-temporal perplexion estimation method through simulation experiments.

In recent years, the sale of products through online auctionsand websites has attracted con-(breakpoint)siderable attention. This studyfocuses on methods for estimating prices from information onwebsites. Conventional estimation methods include statisticalmethods such as regression analysis. However, it is difficult forstatistical methods to estimate prices that include ambiguitiessuch as a seller’s intention or a person by applying rules withina company. Considering these considerations, we postulatedthat fuzzy inference, which can accommodate ambiguous rulesinto account, would prove an efficacious approach. Therefore,this paper examines whether price estimation of goods usingfuzzy inference is effective through simulation experiments.

In recent years, educational support robots have attracted attention, and their educational scope has expanded from school education to driving. Among these, this study focuses on a teacher-type robot that provides correct driving instruction. In driving schools, the GROW model, which provides step-by-step instruction, has been reported to be effective. However, no effective driving behavior instruction method based on the GROW model has been proposed for supervised robots in previous studies. Consequently, this study proposes a driving behavior teaching method for supervised robots based on the GROW model. In this paper, we verify the learning effect of the GROW model-based supervised robot on university students through experiments with subjects.

In recent years, educational support robots have attracted attention. This study focuses on a partner-type robot that learns together while solving problems. Previous studies have indicated that it is possible to improve the learner’s motivation to learn by suggesting the number of problems to be solved by the learner while conversing with the robot before the start of learning. However, the effect of the timing of suggesting the number of problems to be solved, such as before or after learning, on the number of problems solved has not yet been verified. In this paper, we examine the effect on the number of problems solved by learners through two experiments with university students on the timing of the robot’s conversation that promotes the improvement of the number of problems. The experimental results indicated that a comparable number of questions were resolved regardless of the timing of the conversation in which the robot promoted the improvement of the number of problems solved.

In recent years, programming education has become mandatory in elementary, junior high, and high schools, and many educational institutions are using board games to teach programming thinking. The advantages of board games include the ability to learn in a stress-free environment while having fun with others, which can stimulate the desire to learn. However, board games inherently require the presence of an opponent to play with, and it is not always guaranteed that opponent will be available at home. Therefore, this study developed an educational support robot to play programming education board games with students. In this experiment, we investigated the learning effects of playing against this robot on humanities students who had no prior experience with programming. The results showed that competing against the robot provided higher learning effectiveness for humanities students compared to competing against human opponents.

In recent years, with the development of IoT technology, it has become mainstream to manage by electronic data. However, due to the aging of the population in various industries, the number of people who are not good at operating applications or software is increasing. Therefore, a plan that is easy to use even for those who use applications for the first time (hereinafter referred to as beginners) is needed. Therefore, this paper develops an inventory application that performs inventory with a screen agent that assists the operation. The screen agent teaches the operation while explaining how to use it. As a side note, inventory is the work of counting all the products in a company. In the experiment, this application was used for inventory in a company that manages inventory for about 5,000 types of automobile parts every year. The result of this experiment shows that compared to a conventional application that does not use screen agents, using this application could reduce the working time of taking inventory by about 45 percent.

In recent years, there has been a lot of attention paid to educational support robots. In conventional educational support robots, the number of problems to be answered by the learner (hereafter referred to as ”the number of solved problems”) is determined by the system. As a result, the collaborative learning with conventional robots does not create a spontaneous learning environment for the learner. In order to improve this situation, the Problems-Number Negotiation Method has been proposed. It has been reported that a robot using the Problems-Number Negotiation Method improves the number of solved problems during short-term learning with the robot without decreasing the impression of the robot. However, the effect of the impression on long-term learning with a robot using the Problems-Number Negotiation Method has not been verified. In this study, we investigate the effect of long-term learning with a robot using the Problems-Number Negotiation Method on learners’ impression of the robot by conducting an experiment with university students. The experimental results showed that the robot that negotiates the number of questions to be answered through conversation gives the same impression as a conventional robot.

In recent years, as the percentage of children with developmental disabilities in regular classrooms has increased, so has the need for clinical psychologists to administer psychological tests. There are two types of training methods for psychological testing: classroom training and field training. However, due to difficulties in securing training sites and subjects, there is no environment in which clinical psychologists can receive sufficient training in psychological testing. Therefore, in this study, we develop a child-like robot that can train clinical psychologists in intelligence testing (hereinafter referred to as the ”proposed robot”). The proposed robot is equipped with the same speech content as that of a child undergoing intelligence testing. The clinical psychologist trains the robot to perform the intelligence test by talking to the robot. This paper investigated the learning effects of training with a proposed robot on university students studying to become clinical psychologists. The experimental results suggested that training with the proposed robot has the same learning effect as learning to read a conventional manual.

In this paper, we report the results of a preliminary study of diagnosis supporting system of dementia from test images of Trial Making Test(TMT), which is used as a diagnostic test for dementia, by using Data-efficient image Transformer. We verify the usefulness of Data-efficient image Transformer model, which is a highly efficient and accurate data learning method, by comparing with the past-studied model called Vision Transformer.

In bottom trawl fishing, "bycatch fish" has a problem which is wastage of resources to be discarded after being landed. Because of the effective utilization of bycatch fish, Doctor Wada (2021) has research about the food processing to “surimi” from bycatch fish. It’s necessary to identify the species of bycatch fish, because includes both usable and unusable fish species for surimi. However, this previous study identifies the fish species manually with human eyes. Therefore, we proposed to automatically identify bycatch fish species using Convolutional Neural Network. This study verified whether convolutional neural networks can be used to discriminate bycatch fish species. So, we will present the verification conducted in this study.

In recent years, there has been a lot of research on scene recognition, which estimates the situation from a video image. There are two main methods for scene recognition, one using BoVW (Bag of Visual Words) and the other using CNN(Convolutional Neural Network). Both methods use local and visual feature vectors in images to achieve scene recognition. On the other hand, humans recognize scenes and understand the situation by considering the objects they see, their numbers, and the relationships between them. Thus, performing recognition that focuses on the objects in the target image, it is possible to perform scene recognition that semantically understands the situation like humans do. Therefore, we propose an indoor scene recognition method that classifies object information in images obtained by object detection into “static objects ”and “dynamic objects ”and uses the information on the frequency of “dynamic objects ”around “static objects ”as feature vectors.

In recent years, the importance of energy efficiency in computing has increased in the pursuit of a sustainable society. When solving real-world expensive optimization problems, evolutionary algorithms that extensively use iterative computations face challenges of high computational costs and increased energy consumption. This study aims to evaluate how surrogate models can reduce energy consumption to address these issues. Specifically, we applied a particle swarm optimization (PSO) algorithm using a neural network as a surrogate model to a traffic light scheduling problem. This problem requires simulation for solution evaluation, resulting in high evaluation costs. The experimental results demonstrated that a surrogate-assisted PSO can maintain search performance equivalent to a PSO without a surrogate while significantly reducing energy consumption during execution.

When optimization methods are actually applied in industry, it is necessary to capture stake-(breakpoint)holder intervention in mission definition and the resulting changes in problem definition. In this presenta-(breakpoint)tion, I will introduce a case study on the design of a hybrid rocket by applying evolutionary computation to the problem definition using a mission-driven approach, assuming the existence of stakeholders.

The semantic segmentation of bone structures requires pixel-by-pixel classification and high extraction accuracy for objects in the image to build accurate bone models that can be used in diagnoses. Many segmentation methods have been developed, but the most common ones are based on Convolutional Neuronal Networks (CNNs). However, it has been reported that CNN-based segmentation methods cannot extract objects with complex shapes, such as a wrist, with high accuracy. One reason for this is the failure to consider the three-dimensional structure of medical images. Moreover, 3D-CNN methods have been proposed to tackle this problem, but 3D-CNNs require a huge amount of learning data. Considering bone segmentation, we should improve 2D-CNN models to apply to practical uses easily. Therefore, we propose a 2D-CNN-based segmentation method that uses bidirectional convolution processing and reconstructed images to take into account the three-dimensional structure of the bones in the upper limb region. Specifically, our method analyses images from two directions, axial and sagittal, with two different models combining BiConvLSTM and Attention U-Net. The images reconstructed from the sagittal plane to the axial plane are then used to obtain the attention map of the segmentation target. Performance experiments show that the proposed method exhibits an IoU of 0.9355, which is higher than U-Net and the state-of-the-method.

Many methods have been proposed for computer analysis of skin lesions in recent years. Some widely used methods employ rules based on asymmetry, irregularity of boundaries, etc., in skin lesions. However, it is difficult to determine disease from skin lesion images due to their considerable variability, such as noise and low contrast. Therefore, more robust and superior automated skin lesion diagnosis methods are required. Ensemble-based methods have been proposed to improve their robustness, but these methods roughly use the final decision of each model. They can’t effectively use precise image features, which are extracted using each model for the final decisions. In this paper, based on the ISIC open-source dataset, we utilize feature maps from multiple Convolutional Neural Networks（CNN） and construct a Feature Fusion Model by applying Attention to classify skin lesion images into four classes. By applying Grad-CAM to various features extracted from each CNN and concatenating them, we achieved an accuracy of 93%. The proposed model is expected to learn valid features from skin lesion images.

Surrogate-assisted multi-objective evolutionary algorithms (SAMOEAs) are effective ap-(breakpoint)proaches for solving expensive multi-objective optimization problems (EMOPs). However, when the search space of EMOPs becomes high-dimensional, the effectiveness of SAMOEAs significantly degrades. In this work, we investigate how much such performance degradation will occur by comparing SAMOEAs with vanilla evolutionary algorithms. Experimental results show that, surprisingly, SAMOEAs tend to suffer from outperforming vanilla evolutionary algorithms.

We propose the interactive YUKI algorithm as a new approach in interactive evolutionary computation. The YUKI algorithm searches for the optimal solution by balancing convergence and divergence of the current optimal solution candidates. We investigate the performance evaluation of the proposed algorithm through numerical simulations using a pseudo-user model which evaluates the candidate solution instead of real users. The simulation results indicate that, compared to traditional interactive genetic algorithm and interactive tabu search algorithm, the proposed algorithm can effectively balance convergence and diffusion in the search for candidate solutions, resulting in a tendency for higher evolutionary performance when the pseudo user has several preference points.

This study develops and evaluates a method for personal identification using acceleration data. By employing transfer functions between different body parts, the method extracts individual-specific movement characteristics. Specifically, acceleration data from the left thigh to the right pelvis are used, and the transfer functions are calculated using the Z-transform. As a result, transfer functions unique to each individual is computed, enabling personal identification from a set of time-series patterns. This study presents a potentially novel approach to personal identification.

We live with a variety of stresses in our daily lives. One way to release these stresses is to have others listen to our complaints. However, depending on one’s living environment, there may be no one close at hand to listen to one’s complaints. We attempt to solve this problem by using a dialogue system. Tsuboi’s previous approach involved creating a system that responded with interjections and parroting back the user’s statements. This system was examined to see if it could reduce stress. However, it was found that due to the limited response patterns and content, the system felt impersonal. In this paper, we propose to enhance Tsuboi’s system by increasing the variety of interjections and responses, and by adding methods for emotional expression based on emotional analysis.

The purpose of this research is to encourage users to engage with chat dialogue systems by making them feel more familiar and natural during interactions. To achieve this, we propose a chat dialogue system that recognizes the characteristics of user utterances and generates responses with similar features. Three elements are considered as the characteristics of user utterances: habitual expressions, politeness, and personal pronouns. By fine-tuning a T5 model on a corpus created from Twitter (X) data, we construct a model that generates response sentences with specified characteristics (habitual expressions and politeness). Furthermore, post-processing is applied to adjust the system's responses to match the personal pronouns used by the user. Evaluation experiments are conducted to investigate the effects of the system's responses, tailored to users' linguistic characteristics, on user engagement and satisfaction.

In this paper, we apply the Deep Knowledge Tracing (DKT) method to estimate a user’s knowledge state using both the problem-side knowledge model and the user-side learning process model in learner response simulations. Our goal is to confirm changes in the learner’s knowledge state concern-(breakpoint)ing problem progression and structure. Specifically, we conduct learner response simulations based on a knowledge keyword model of problems, which has been previously organized using keywords. We compare our approach with a simple word-based response simulation as a baseline. The results confirm that DKT is effective in observing changes in the learner’s knowledge state when applied to these simulations.

In recent years, research on in-home behavior recognition has been active against the backdrop of the spread of smart homes. One of the challenges in this field is the time-consuming process of collecting sensor data installed in the home for a certain period of time and annotating the data in order to construct a model for in-home behavior recognition. In response to this, methods to reduce the cost of sensor data collection and annotation by transfer learning have attracted attention. However, a lack of datasets for experiments is an issue when considering the adaptation of transfer learning to in-home behavior recognition. A common method to create a dataset is to measure real data, but this is expensive to prepare a residential facility for the experiment. In addition, it is difficult to create data sets under various conditions because subjects are held for a long period of time. The purpose of this study is to facilitate the creation of datasets by simulating in-home behavior and generating simulated datasets. Using the generated dataset, we will conduct experiments on the application of transfer learning to in-home behavior recognition.

In the early stage of BMI research, studies attempted to classify natural thoughts such as “ I want to go left ” or “ I want to go right ” based on EEG signals, but the classification accuracy did not improve. Consequently, to enhance accuracy, more classifiable yet unnatural thought patterns were employed. However, controlling BMI with unnatural thoughts is challenging. Moreover, current EEG measurement devices and analytical methods have progressed since the early BMI research days. Therefore, this study aims to verify the feasibility of bidirectional classification based on natural thoughts using various analytical methods, including deep learning techniques.

The Kestenberg Movement Profile (KMP) is a method of movement analysis based on a psychoanalytic perspective. Usually, in KMP, the analyst creates a recorded waveform, called a rhythm line (notation), from the subject's natural daily movements. The analysis is then performed by classifying the basic patterns (pure rhythms) and making relative shape (attributes) judgments. However, since the rhythm lines are drawn based on the physical empathy of the analyst, there is ambiguity in the obtained rhythm lines, such as the analyst's senses and habits. Therefore, the field of clinical application calls for the development of a system that automatically analyzes rhythm lines. In our previous studies, we developed a pen tablet-based rhythm line input system and examined feature extraction for extracting pure rhythms. This paper reports on the development of a tool to support feature determination of rhythm lines using the proposed method (Fuzzy Sets).

The inner class model, a model of general artificial intelligence, is hereby defined as a type of recurrent neural network. Although the inner class model is a type of reinforcement learning, its learning does not depend on rewards, but depends on how often the strengthening of knowledge is interfered by noise. Here, to strengthen the knowledge is to repeat the time-series episode after the episode took place with action. In addition, more noise supposes to be input from outside when an action fails than when it succeeds, and it referred to as "noise hypothesis." In the new definition, generating an inner class is equivalent to form a node cluster, and node clusters are formed by slightly increasing the weight of edges every time an edge is propagated. It was confirmed by a simple time series episode and by a small-scale RNN that node clusters are created by this new definition and that it is possible to repeat the episode.

The National Association of Commercial High Schools, a public interest incorporated foundation established with the aim of contributing to the development and improvement of commercial education in high schools, conducts a variety of qualification examinations. In this study, one of these tests, the “English Proficiency Test” is focused on and examined in terms of metrical linguistics. In short, frequency characteristics of character- and word-appearance are investigated using a program written in C++. These characteristics are approximated by an exponential function. Furthermore, the percentage of Japanese junior high school required vocabulary and American basic vocabulary is calculated to obtain the difficulty-level of each material.

In recent years, business katakana words such as “internship” and “compliance” have been used frequently. However, the reasons why business katakana words are used more frequently and the line between common and uncommon business katakana words have not been examined until now. In this study, we examined the characteristics of business katakana words and the evolution of their meanings by morphological analysis of newspaper big data using word2vec. Based on the results of the morphological analysis, we grouped the business katakana words and examined “business katakana words that should be avoided” and “business katakana words that should not be avoided”.

This study evaluates the impact of dataset characteristics and annotation biases on argumen-(breakpoint)tative relations. Since datasets annotate actual discussions, the proportion of linked arguments is low, and the frequency varies depending on the type of relation. Additionally, direct comparison is challenging due to the differing goals of the dialogues and the individuality of annotators for each corpus. This research conducts estimations of argumentative relations under various conditions and reassesses the validity of existing corpora to explore methods for improving argument analysis.

The purpose of this study is to conduct a text analysis of local assembly minutes to compare and analyze regional issues, concerns, and approaches. The target areas of this study are the 19 municipalities within the Tokachi Subprefecture, based on the administrative divisions of Hokkaido. The study will investigate the availability of minutes available on the website for these regions and, using the collected minutes, perform text analysis to compare specific issues.

In this paper, we test the effectiveness of a method for role-based classification of statements in the National Diet using a BERT-based classifier. Experimental results using two models, FP_MinWiki, which was pre-trained with the Local Meeting Minutes, the National Diet Record, and Wikipedia, and the Tohoku University Japanese BERT model, which was pre-trained with a wide range of Japanese data including Wikipedia, showed that the accuracy of the "opinion/non-opinion" binary classification experiment was The accuracy of the "opinion/non-opinion" binary classification experiment was 87.08% for FP_MinWiki and 74.72% for the Tohoku University model. These results indicate that even training data of about 720 sentences can be classified with more than 80% accuracy in binary classification, and that the characteristics of the data used for pre-training are reflected in the classification results.

The purpose of this study is to investigate the effectiveness of child-rearing support programs administered by local governments, utilizing an information website related to pregnancy, childbirth, and child-rearing. Specifically, we will examine child-rearing support by correlating the names of local government-run child-(breakpoint)rearing support programs, comments made during local assembly discussions on child-rearing and posts on the information website related to pregnancy, childbirth, and child-rearing.

This study aims to analyze the texts exchanged in online communities where pregnant and parenting women share information, with the goal of early detection of physical and mental health issues in mothers and children, and timely provision of relevant information. It is believed that search behavior in online communities reflects the anxieties and questions of mothers. Onomatopoeia, in particular, is useful for expressing pain, symptoms, and emotions, and is closely related to mothers’ anxieties. This paper focuses on the onomatopoeia included in search keywords and attempts to visualize specific concerns of mothers by clustering the words searched together with onomatopoeia.

Federated learning is a scheme of analyzing distributed data by preserving personal privacy and has been applied to fuzzy clustering. In this paper, federated learning-based linear fuzzy clustering is applied to a horizontally distributed sensor data, where characteristics of the federated learning model is discussed through comparison with the batch learning result.

The performance of collaborative filtering (CF) can be improved by utilizing not only user-item cooccurrence information but also additional information. This paper improves co-clustering-based CF by introducing three-mode fuzzy co-clustering, which utilizes the conventional user-item cooccurrence information in conjunction with additional genre information on each item. User-item co-clusters are extracted so that preference tendencies of users on items are considered with their intrinsic preferences on genre categories. Then, the recommendation capability of co-clustering-based CF is expected to be improved even when cooccurrence information is quite sparse. In numerical experiments with MovieLens benchmark data, recommendation performance is demonstrated to be improved by properly increasing the responsibility degree of genre information.

Network data consists of vertex nodes and edges connecting nodes. Network data can represent various real-world events as networks, such as SNS follow-up relationships and railway lines. Analysing network data is important in that it provides insights that take into account not only information about nodes, but also the relationships between them. However, existing network analysis methods have issues such as high computational cost and a limited number of analysis methods. Node embedding is a method that embeds network data into a low-dimensional vector space. This allows network data to be treated as vector data and enables visualization and classification using machine learning methods. In this study, we focus on network clustering using node embedding. Experimental results show that network clustering with node embedding outperforms existing methods in some cases. This suggests the usefulness of clustering using node embedding as a new clustering method for network data.

ICCOECBON (Intra-Cluster COntrolled Edge-sized Clustering Based on Optimisation for Network data) is a clustering method for network data. ICCOECBON++ is also a method that introduces the initial value determination approach of k-means++ into ICCOECBON. In our previous studies, diffusion kernel was used to calculate dissimilarities. Diffusion kernel has the problem that kernel parameters need to be set and cluster partitions depends on the kernel parameters. In this study, structural similarity, which does not require parameters, is used to compare performance with the diffusion kernel. Four benchmark datasets were used in the experiments to compare with conventional methods. The experiments suggest that structural similarity gives better results than diffusion kernel.

Attribute reduction is a data analysis method based on rough set theory. It is defined by preserving various measures and/or structures derived from rough set models. In this study, we propose attribute reduction based on purity measures in the probabilistic rough set model. Purity measure is a measure of the uncertainty associated with objects, whether they belong to a target set. We investigate the relationship between conventional attribute reduction methods and purity-based attribute reduction, as well as clarify the relationship between the convexity of purity measures and the monotonicity of reducts.

We have investigated rule generation under uncertain information and proposed the DIS-Apriori and the NIS-Apriori algorithms. Using such algorithms and the implemented software tools, we can handle rules. This paper considers detecting dependency between attributes based on the obtained rules. Intuitively, the sum of the support values of the obtained rules becomes the degree of dependency. Therefore, we try to find attributes of rules with high support values. Such attributes are candidates that derive dependency.

When extracting any information from an information table with incomplete information, following Lipski we only know the lower and upper bounds of the information. Methods of rough sets that are applied to data tables containing incomplete information are examined from the viewpoint of Lipski’s approach based on possible world semantics. It is clarified that the formula that is first used by Kryszkiewicz, which most of the authors use, only gives the lower bound of the lower approximation and the upper bound of the upper approximation. This is due to that the formula considers only the indiscernibility of missing values with another value. We extend Kryszkiewicz’s formula by considering the discernibility of missing values. As a result, the extended Kryszkiewicz’s formula gives the same approximations as those in terms of Lipski’s approach.

We tried to classify the sleep stages of polysomnographic data using deep learning. The Sleep Heart Health Study (SHHS) dataset published by the National Sleep Research Resource (NSRR) was used to train and classify the convolutional neural network (CNN) model. We report on our evaluation and discussion of the results of using a CNN model to classify polysomnographic data into six sleep stages.

Data science is being increasingly emphasized in university education, and student are required to develop the ability to handle data correctly and address complex real-world challenges. In recent years, the use of tools such as generative AI has created an environment where even beginners can easily perform data analysis. Among these techniques, cluster analysis, which can reveal data patterns by grouping similar data points, is widely used. However, the choice of appropriate data preprocessing methods is difficult for beginners, as the results can vary significantly depending on the preprocessing method used. Therefore, this research focuses on the relationship between data preprocessing patterns and analysis results in cluster analysis. By clarifying the characteristics of this relationship, we aim to support beginners in selecting appropriate preprocessing methods. Specifically, we will use financial data from 80 listed companies, apply four different preprocessing patterns, and then perform cluster analysis to identify the differences in results and characteristics that emerge.

Electron Back Scattered Diffraction Pattern (EBSD) is commonly used to evaluate the crys-(breakpoint)tallinity of metals. However, the equipment used for EBSD measurements is expensive, and only a limited number of institutions own such equipment, making the measurement time-consuming and costly. There-(breakpoint)fore, we have been investigating a simpler and quicker method to evaluate the crystallinity of metals. We used machine learning to correlate the solidification pattern of liquid gallium with the Euler angle mea-(breakpoint)sured by EBSD to predict the crystallographic orientation of gallium.

As an easy-to-use approach to safe driving support, we have developed driver status monitoring systems using wrist-worn accelerometers. In our previous research, we proposed an onboard measurement system to collect driving data in real vehicles and conducted measurement experiments. In this paper, we discuss the results of the experiments and our approach to estimating driver activities from the wrist acceleration data collected in real vehicles.

In this study, based on the structure of visual information processing, in the brain. We propose a recurrent CNN model that has top-down processing and recursive processing within the same layer. We evaluate the effectiveness of the proposed model in the tasks of image classification, video classification, and video prediction. Our model may contribute to robustness in image classification, but it is not superior to existing models in video classification and video prediction.

In this study, we propose an action selection probability model that takes into account the case where there is a certain bias in the action to be selected in reinforcement learning-based action modeling. In the proposed action selection probability model, the softmax function is shifted in parallel when calculating the action selection probability, assuming that factors other than reward influence the selection of actions. Specifically, parallel shift is achieved by adding a certain bias to the difference of action values in each state for calculating the action selection probability. In the proposed method, this bias value is determined based on maximum likelihood estimation in addition to the learning rate and inverse temperature in conventional reinforcement learning models, respectively. In order to confirm the effectiveness of the proposed method, we artificially generated data that is likely to take a certain action independent of the reward using a two-armed bandit problem, which is a type of benchmarking, and compared the likelihood of each model in the conventional and proposed methods using this data. The results showed that the likelihood of the proposed method was significantly higher than that of the conventional method.

In recent years, research on ”artificial life” which emulates behavior based on the biological principles of organisms in semi-artificial living systems has gained significant attention. This study im-(breakpoint)plemented a hardware hybrid agent with a voice I/O interface that connects to the outside world. The agent uses a living neuronal network of cultured dissociated neurons as its core to reflect the‘ fluctuation’ of biological signals in behavior. The agent converts voice input to text, analyzes its sentiment using a Japanese sentiment dictionary, and computes an emotional polarity value. Based on this value, constant current stimulation is applied to the cultured neuronal network. The neuronal activity patterns induced by the stimulation are identified by learning-type fuzzy template matching (FTM), and the agent expresses the mood of the biological neuronal network through pitch. The mood is defined by the distance between pre-made templates corresponding to pleasant/unpleasant responses and the input pattern, expressed in the pitch of the output sound. When speech is directed at the agent, the output response reflects the history of all inputs, showing a tendency to express a summary of continuous pleasant/unpleasant reac-(breakpoint)tions rather than a single reaction. Furthermore, depending on the spontaneous activity of the core living neuronal network, various individual tendencies, such as overall pleasant or unpleasant emotional trends, were observed.

Under the current research, it has not yet been revealed how various emotions are expressed in the brain under verbal stimuli. In order to clarify this point, we will use fMRI Datasets, which contains BOLD responses measured by functional MRI while human subjects listen to a reading as a language stimulus, to investigate how sentences given as a verbal stimulus affect emotions in the human brain. First, in order to obtain the emotion ratings for each sentence, we placed 80 annotators for each emotion and obtained the emotion ratings. Next, we fine-tuned the pre-trained BERT with GoEmotions, then fine-tuned it with another 80 emotion ratings, and obtained the emotion features for each sentence. Then, based on the model obtained by regression of the concatenated language and emotion features with the measured brain activity, we retrieved the predictive brain activity for each emotion, and investigated and discussed how the predictive brain activity responds to each emotion.

In recent years, research utilizing EEG for emotion recognition has garnered significant attention. Previous studies have shown that deep learning methods can recognize pleasant and unpleasant emotions triggered by memory with relatively high accuracy. However, the methods currently considered to have high recognition accuracy relay on multi-channel brainwave measurement, big data, which are costly. Therefore, research that ensures accuracy depending on a small amount of learning data is crucial for technological advancement in this field. In this study, we propose an emotion recognition using “fuzzy deep learning.” His method involves classifying converted fuzzy feature images through transfer learning with a pre-trained VGG16 image classification model. We integrated the features of small-channel brainwave data in a fuzzy inference framework and converts them into images through fuzzification using Learning-type Fuzzy Template Matching (L-FTM). The generation mechanism of the fuzzy feature image provides an interface that integrates brainwaves with different characteristics such as signal/noise ratio, average amplitude, and time series changes, without preprocessing. This approach allows the use of individual biometric data for learning.

Voice recognition technology allows for the control of external devices without the need for physical movements. However, it can sometimes misinterpret ambient sounds or respond to unintentionally spoken startup commands. Triggering action not intended by the user. Therefore, research is underway to decern the“ user’s intention ”at the time of voice command activation. In this study, we attempt to identify user intentions using Brain-Computer Interface (BCI), which learns from voice and EEG data though a Learning-type Fuzzy Template Matching(L-FTM) method. The L-FTM method employs linguistic labels such as ’High’ and ’Low’ instead of actual values to ambiguously represent the characteristics of the input data. This approach allows the L-FTM method to learn the features of input data while suppressing the effects of noise in data such as EEG, which is known to have significant individual differences, and speech data that includes environmental sounds and noises. Participants were seated in a quiet room, in front of a desk wearing a head cap connected to an electroencephalograph. The desk was equipped with a PC for stimulus presentation and audio acquisition, and a Smart Speaker. During the leaning phases, buttons labeled,“ with intention ”and“ without intention ”were displayed on the PC screen. Participants were asked to vocalize “ Alexa ” after selecting either button, and the vocalizations with and without intention were recorded 30 times each. When determining the presence or absence of command intention, participants were asked to perform the same task as during the learning phase, with six vocalizations of “ with intention ” and six vocalizations of “ without intention ” each, Subsequently participants were asked to identify whether these vocalizations were intentional or not.

In recent years, the importance of a speech segment detection technique called speaker di-(breakpoint)arization has increased, mainly in conferences, news, and telephone calls. However, conventional speaker segmentation detection methods using neural networks require a huge amount of training data. In this study, training data was created by recording the speech of two speakers of the same gender, splitting and combining them to create a synthetic utterance. The effect of the distance and angle to the microphone on the accuracy of the test data was examined in tests with non-synthesized speech.