In recent years, With the labor shortage due to the declining birthrate and aging population, the demand for service robots that provide information support such as guidance and recommendation is expected to increase furthermore. However, it is still difficult for service robots to deal with ambiguous requests, such as uncertain or unclear requests of individual users. In this paper, we propose to provide integrated information recommendation tailored to users through interaction with users. The purpose of the system is to provide information support according to the estimation of the user's potential intentions by introducing the reliability estimation system with user reactions as input.

In this study, we investigate whether a human-robot scenario continuously had influence on participants after the scenario. Many studies on the social behaviors of robots have been conducted. It is important that these robots try to naturally participate in a human community and behave in a human-like way. As robots get sociable, humans that interact with the robots are likely to be affected by the robots that behave in a human-like manner like they are affected by other humans. In particular, some studies show that robots had influence on humans in some human-robot experimental scenarios. Although several previous studies about social robots investigated the social influence on a human from robots in the human-robot scenario, long-lasting influence on a human after the scenario still incompletely understood. In this study, we investigate the long-lasting effect on human decision-making in an experimental scenario of human-robot groups, which included robots learning group norms. We assess this influence by analyzing the results of two kind of questionnaires that the participants answered during the experimental human-robot scenario and more than one week after the scenario. The questionnaire results reveal that the decision-makings of some participants were limited by a group norm developed in a human-robot group more than one week after the experimental scenario.

In this study, we developed an autonomous communication system using a humanoid robot and investigate the effect of the robot’s movements on the impressions and reactions of people during the blank in the conversation between a robot and a person. In this paper, we construct an experiment for this study and describe the results of the preliminary experiments.

In the field of pet robots and robot-assisted therapy (RAT), characterization of animal motion is important for the development of robots resembling various animals. This paper presents a method for the generation of animal gait in quadrupedal robots. In this study, we employed AIBO as an experimental quadrupedal robot and generated the gait of the robot on the basis of an animal’s gait and zoology. Moreover, we realized the stable gait on the ground by adjusting the minor deviation of parameters for each joint using chaotic time series analysis.

In recent years, researches on image inpainting by deep learning are rapidly progressing. Until now, there have been many studies on partial inpainting of facial images and landscape images with Generative Adversarial Network (GAN). But it was difficult to apply it in real time. In this research, a GAN-based image inpainting method that can reduce learning and processing time is proposed. In this method, GAN is used for the network structure, and the mask is updated using the Gaussian filter in the generator part of GAN. Through verification experiments using actual images, this method is several to 10 times faster than conventional methods such as Partial Convolution and Gated Convolution.

Many studies on handwritten figure recognition have been published so far. For handwritten graphics in the fields of science and engineering, relatively clear drawing rules are defined, such as the symbols of the constituent elements being predetermined. On the other hand, there is no particular drawing rule for sketches such as landscape paintings, which makes it difficult to design a feature extractor. Against this background, in recent years, research on recognition models using deep learning from sketch images such as landscape paintings has been reported. However, the recognition rate of all the research results is about 70% or less. Therefore, in this paper, we proposed a new CNN model for recognizing landscape sketch images, and verified its effectiveness by computer experiments. Since there is no benchmark data for landscape sketch images, it cannot be compared with the results of previous studies, but the recognition rate of the CNN model proposed in the text is about 80%, which is higher than that of previous studies.

In this paper, we examine the improvement of concentration and inactivity through exercise in an office work situation. The purpose of this is to divide long hours of work into short periods of time, and to exercise in between, so that the lack of exercise and concentration could be improved. Evaluation was made by analyzing the percentage of correct answers to the calibration task and biometric information. The results of the experiment showed no significant trend in biometric information, however there was an increasing trend in the percentage of correct answers for the condition with exercise in the two combinations of the two conditions. Therefore, it is suggested that divided into short periods of time and doing exercise when work that takes a long time may help you to maintain a higher level of concentration than with continuous work.

The websites of local governments have information about various administrative services. It is considered that many local residents make actual applications by looking at the information on these sites. Therefore, not only the visibility and operability of the website itself, but also the accessibility to necessary information is important. On the other hand, it is considered that the evaluation points to be emphasized are different for each website user. In this study, we aimed to evaluate the usability of the home pages of local governments by considering individual judgment criteria. Also, we tried to evaluate the websites of local governments using fuzzy measures.

Convolutional Neural Networks (CNNs) achieve high classification performance in object recognition. Although CNNs automatically learn their network weights by data-driven methods, the appropriate hyperparameters of the network structure must be selected to achieve high classification performance. In general, the appropriate hyperparameters depend on the dataset to be used. In addition, because finding the optimal hyperparameters by manual adjustment is difficult, automatic adjustment using search methods has been actively studied. In this research, we optimize the number of filters, which is one of the hyperparameters in CNNs, by using evolutionary multi-objective optimization to enhance the generalization ability. Specifically, we use multiple datasets and define multiple objective functions based on the loss functions for those datasets. That is, the number of objectives is the same as the number of datasets. In computational experiments, we compare the classification accuracy of CNNs optimized by the multi-objective optimization method and the single-objective optimization method.

In some real-world data mining applications, multiple class labels are assigned to a single pattern. Classification problems with these patterns are said to be multi-label classification. Recently, some interpretable classifier design algorithms have been proposed for multi-label classification. Multi-objective fuzzy genetics-based machine learning is one of the most interpretable classifier design algorithms. This algorithm optimizes a number of fuzzy rule-based classifiers based on maximizing accuracy and minimizing complexity. Because there are several accuracy measures for multi-label classification, multiple runs with a different measure are necessary when we want to obtain the best classifier in terms of each of those accuracy measures. In this paper, we compare three two-objective formulations (i.e., one accuracy measure and one complexity measure) and a four-objective formulation (i.e., three accuracy measures and one complexity measure) in fuzzy genetics-based machine learning for multi-label classification.

Multi-modal multi-objective evolutionary algorithms can solve a problem with multiple Pareto optimal solutions that have the same objective function vector. Most of existing multi-modal multi-objective evolutionary algorithms use the convergence in the objective space as the primarily fitness evaluation criterion. As a result, they do not always have high approximation ability of the Pareto set in the decision space. To approximate better both the Pareto front and the Pareto set, we propose a multi-modal multi-objective evolutionary algorithm based on problem transformation. A multiobjective optimization problem is transformed into a number of two-objective subproblems. In each subproblem, solutions are optimized in terms of the corresponding scalarizing function and the decision space diversity. The proposed algorithm can maintain not only solutions with good convergence to the Pareto front but also diverse solutions in the decision space. Experimental results show that the proposed algorithm has a high approximation ability to both the Pareto front and the Pareto set.

We previously proposed the evolutionary algorithm inspired by the evolution of birdsong grammars. One of the objectives for the proposal was to quickly obtain practical solutions to large-scaled problems. However, we did not apply the algorithm to large-scaled problems there. So, in the paper, we apply the algorithm to a large-scaled OneMax problem and its variants. The simulation results show that the algorithm is likely to produce solutions with simple bit patters and with not so bad fitness values for large-scaled problems as a bit pattern of the global optimum becomes more random. However, when the problem size is smaller, the algorithm is likely to produce a precise bit pattern of the global optimum.

In this paper, we propose real-coded genetic algorithms that utilize a method for detecting dependency relationships between variables. The method consists of neural network regression and group lasso. The proposed genetic algorithms select an appropriate crossover operator based on the dependency information between the variables, which are obtained from past solution candidates. Simulation results using the CEC’13 benchmark functions show that the proposed algorithms outperform conventional real-coded genetic algorithms.

Wearable biometric sensors are generally expensive.Therefore, this study uses an Arduino and a low-cost version of a pulse sensor to inexpensively measure biological signals.

In recent years, although fatal traffic accidents have been decreasing, there are still many accidents due to drowsy driving. The sleepiness detection of a driver has been using the physiological indexes. On the other hand, in the fatigue study, the sleepiness is related to the subsidiary behaviors. In this paper, we report on the result of the sleepiness detection focused on the driveŕs action characteristic.

We examine a possibility of personal authentication using forearm surface EMG (s-EMG) during gesture operation. The s-EMG was measured 10 times from the five subjects during performing the 6 gestures. In our previous research, we performed the gesture identiflcation by using SVM. As the feature values, we use the maximum and minimum values and their times of the time domain data during gesture. As the result, the identiflcation rate was 66.7%. In this paper, to improve the identiflcation rate, we increase the future values which we use, and we introduce the selection algorithms of the important feature values, ANOVA and random forest. As the result, the identiflcation rate is improved to over 70% with both selection algorithms. Moreover, the previous research claimed that the feature values in frequency domain was not effective. However, we found that some feature values in frequency domain was effective.

This study devises a Bayesian network model to infer the occurrence of difficulty in character identification when browsing a website. Subjects are classified into two groups: one is a group of people with fast eye movement and the other is that of people with slow eye movement. It is shown that, in either of groups, the average fixation duration when feeling difficulty in character identification is significantly longer than when feeling difficulty in sentence comprehension. Browsing properties of subjects in both groups are analyzed by inputting time series data consisting of eye movement values in two previous periods into the inference model. It turns out that in both groups, the occurrence probability of the difficulty in character identification tends to increase when subjects move their eyes rapidly to the character from a distant position or move their eyes closer to the character little by little.

In this study, English sentences of several private-sector English tests that the Japan’s education ministry has been considering to introduce as part of the university entrance common test are 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.

Many local governments regularly survey their residents for administration evaluation and publish the results in their web pages. When we compare those results, we can find the characteristics of each local government. However, since each questionnaire consists of its own questions even if the issue is similar, it is difficult to associate one question by the local government to that by the other. We propose the system to classify the questions considering word order and meaning by Long Short Term Memory (LSTM) with word embedding.

In this paper, we report a case study of constructing a machine learning model from real data. The case report in this paper is to determine whether a child needs school support or not. This case study uses a Yamano-style screening sheet to determine the support for elementary school children. Conventionally, the decision on whether or not support is needed was made by a person. However, when the number of children increases, the amount of support exceeds the tractable amount. In addition, there was a problem that it was difficult to obtain a common standard of support-decision because of the total dependence of such decisions on local governments and schools. Therefore, a support-decision model was created by machine learning from screening data that were collected from the elementary schools. The procedures used to construct the model from real data, the accuracy of the model obtained, the lessons learned from the series of work, and future issues are presented.

The authors have defined a type-2 fuzzy contingency table for contingency analysis and proposed a fuzzy portfolio analysis method based on it. In the conventional fuzzy contingency table, the intersection of fuzzy sets is defined by min operation, but we replace it with t-norm and propose a new type-1 fuzzy contingency table. In this paper, we redefine fuzzy contingency table using t-norm, and furthermore, we introduce fuzzy portfolio analysis as an application of the technique.

In this paper, we focus on multiobjective two-level simple recourse programming problems with discrete-type LR fuzzy random variables. To deal with such problems, we introduce new solution concepts called optimistic and pessimistic Pareto Stackelberg solutions for the leader. It is shown that such optimistic and pessimistic Pareto Stackelberg solutions can be obtained by solving weighting problems of the leader. We propose an interactive algorithm to obtain a satisfactory solution of the leader from among an optimistic or pessimistic Pareto Stackelberg solution set.

In this paper, we propose an analysis method with fuzzy inference in order to improve the classification accuracy of team strategies in RoboCup soccer. It is currently difficult to quantitatively evaluate team strategies because there are no appropriate ways to represent game situations and also there are an intractable number of factors such as field states and tactics. Therefore, the performance of tactical analysis is not high enough to identify unknown teams. Because the kick probability distribution proposed in the previous works cannot consider the kick directions, this paper employs a kick direction distribution obtained by kernel density estimation using von-Mises distributions. In a series of computational experiments, a fuzzy inference system with the kick probability distribution as well as kick direction distribution in the antecedent is constructed from the RoboCup games. This paper evaluates the classification accuracy in order to investigate the performance of the proposed method.

Although the automation of inspection processes for various kinds of industrial products has progressed, the situation seems to be largely depending on visual inspection ability of inspectors who are familiar with the quality control of each product. Recently, not a few attempts have been tried to apply convolutional neural networks (CNNs) specialized in deep learning technology to image recognition for product defect detection. The authors have developed an application that can design and train CNNs and support vector machines (SVMs). In this paper, the application is tried to be applied to the defect detection in the manufacturing process of wrap roll product. Firstly, a template matching technique is used to extract only the target film areas from the entire images of wrap roll products. Next, a CNN named sssNet consisting of 15 layers is originally designed so as to classify input images into defective or not, then trained using a large number of original and augmented images to enhance the generalization ability. Finally, the trained sssNet is evaluated through classification experiments of test images. The usefulness of the developed application with a promising function of defect visualization is also assessed through this test trial.

In recent years, multiple people work at same place. In those places, workers are disturbed by some noises. This paper researches the audio-visual control system which encourage concentration in those places. In real working space, there is three types of information. First type is the main task information which is worked mainly. Second type is the sub task information which people have to notice. For example, it is notification and visitors. Third type is the noise. It is unrelated information. In order to encourage concentration, the noise should be blocked, tasks should be reinforced and information which improve the concentration should be added. This paper aims to make prototype system which have the noise blocking function and survey effect of this prototype. This system has video see through HMD and hear through earphone. In the experiment, subjects are imposed a simple calculation task for the main task and mind random video change for the sub task, which are experimented with and without the system function. The main task results don’t show a significant difference. But the sub task results show a significant difference. The sub task results with function is faster than them without function. The system function makes easier for user to notice sub task. In the future, the prototype system will be added the functions for reinforcement and improvement and be experimented in real working environment.

In recent years, virtual reality (VR) has been used in various fields such as entertainment, medical care, education, and training, but the use of VR causes a problem of gorilla arm syndrome. This is a problem caused by fatigue, which hinders comfortable aerial interaction in VR. Therefore, we flrst create a simulator that analyzes the accumulation of fatigue that causes gorilla arm syndrome. Using this simulator, we aim to improve the gorilla arm syndrome by verifying an appropriate support method to reduce fatigue in the main movements performed in VR.

Due to the structure of the HMD (Head Mounted Display), the resolution of the peripheral vision is low and it is difficult to read the text in the peripheral vision during text reading. In this study, we design a VR (virtual reality) text reader that improves readability by dynamically controlling the size of characters and increasing the size of characters away from the central vision.

It has already been proved that WIP (walking-in-space) is used to reduce the effects of VR sickness. Although WIP systems provide a sense of movement, it is easily to cause fatigue that users always lift their legs up and down. Well in this paper we want to find out whether it is effective to reduces VR sickness by cycling exercise on both legs like a WIP when riding an exercise bike. In addition, we verify whether the bike can reduce VR sickness by moving legs with the motor within the exercise bike and by this way to reduce the fatigue caused by WIP.

Characteristic of the optical see-through HMD is that the characters and images to be displayed are transparent and the other side can be seen. Due to the characteristics, the display of the HMD may be difficult to see depending on the real space (outside world) that is the background of the display. Because of that problem, in addition to research on increasing the opacity of HMD display, research on changing the transmittance of the external environment, which is the background of HMD display, is being conducted. Then, by combining them, we decided to investigate when each condition makes the HMD display easier to see. However, it is difficult to evaluate the actual optical see-through HMD by changing the display and the transmittance of the outside world. Therefore, we decided to investigate the optimal combination of these by simulating an optical see-through HMD in a VR environment, and changing the transmittance of the HMD display and the external environment in it.

The latest technologies and interactions in the field of virtual reality (VR) have brought progress to VR locomotion. The VR locomotion methods that have been studied so far are diverse, including real movement, WIP, teleportation, joystick-based, and gesture-based. However, these movement methods are not dealt with in this study. These are two-dimensional movement methods in which the user moves on the ground and is not possible without the ground. As a result of several years of research, it has made remarkable progress and has created an innovative transportation method called "teleportation". In this research, we propose a three-dimensional moving method that moves in the air instead of only on the ground, which is different from the conventional moving method. Furthermore, we propose an interaction that utilizes the line of sight and the inclination of the head, which is an extension of teleportation.

In this paper, we develop the posture control system for multi-legged walking robot HEXA. In this system, we attach depth camera to HEXA and get information of obstacles and passages. In order to achieve HEXA's smooth moving, we develop the program to adjust the height of HEXA's body and walking mode based on the information.

This paper develops a control scheme for electric prosthetic hand system based on camera images. We develop a simulation environment of an electric prosthetic hand based on camera images by Unity, and acquire control scheme by reinforcement learning in the simulation environment.

We propose a You Only Look Once (YOLO) based road marking detection method for self- position estimation in autonomous cars. In the proposed method, YOLO pre-processing is added to a conventional detection method, which searches road marking from a whole input image including various background, to restrict the searching area by removing background areas. In the conventional method, it is necessary to set a high detection threshold to prevent false detections. On the other hand, the proposed method can set a lower threshold than the conventional method because the searching area is restricted to road marking areas only. Therefore, several road markings that were previously undetected can be detected. The experimental results show that the proposed method reduces the number of negative detection while keeping the number of false detections at zero. Also, the overall accuracy was improved by 1.3 points.

The development of robotics has dramatically improved the robot's ability to recognize the environment. However, the method by which robots estimate human emotions is difficult to solve. In this paper, we propose a method to estimate human emotions in real time from human facial expressions, spoken words, and voices.

Table talk Roll Playing Game is a conversational social communication game in which participants with different 2 roles progress by conversation. The remarks made by Game Master, the host of TRPG, affect the fun of the game. The ambiguous speaking styles that GM frequently uses in the game can be expected to have the effect of prompting the participants to think and helping the game progress. In addition, since we believe that ambiguous speech assists deepening the relationship between humans and systems in a positive direction, this research investigates ambiguous speech in systems. As the first step, we develop a bot that plays a role of a game master in TRPG, and investigate and verify the way of speaking that gives a good impression to the user. We developed and used two types of GM bots, which are control conditions that use only affirmative forms, and GM bots, which are ambiguous conditions that use ambiguous endings such as “maybe” to imply. We asked experiment participant to use one of the GM Bot and evaluated it using the Impression evaluation scale in CMC. As a result of the evaluation, the values of "friendly type" and "feel a sense of trust" Gained Significantly higher in the ambiguous GM Bot. In addition, a GM utterance model was created, and comparative experiments were conducted using model adapted and unadapted TRPG playlogs. The subsequent research policy will be examined.

Recently, research and development regarding education support robots has been conducted. In this study, we focus on a partner-type robot, such as a collaborative learner. According to a previous study, human learners lose likeability toward robots as learning advances, and thus, cannot perform collaborative learning with them. To address this problem, we must maintain the learner's likeability toward the robot. It is important to make the learner realize that they learn more with a robot. Therefore, this study develops a partner-type robot that learns by observing the learner. The proposed robot solves problems along with the learner in an alternating manner. When the learner solves the problem correctly, the robot learns the answer. Moreover, when the robot answers a question, it asks the learner whether the solution is correct. Thus, the robot emphasizes to the learner that it does observational learning. In addition, the robot makes the learner strongly realize that they learn with it. Thus, the robot can promote the learner's likeability toward itself. This study investigates the impression evaluation that a robot that learns by observing a learner gives to the learner. In impression evaluation, we evaluate likeability and perceived intelligence of the robot. The experimental result shows that our robot can improve the learner’s likeability with itself via collaborative learning.

In this study, we aim to develop an educational support system merging information technology (IT), network technology (NT), and robot technology (RT). However, it is difficult for teachers and teaching staffs to extract essential clues from huge amount of data to give appropriate feedback to students. Intelligent technology is therefore required for the data analysis and feature extraction. The heart rate variability was measured during the equation problem task on the display with yellow and blue colored background images. From the psychological and physiological analysis, we found that yellow color was the more efficient background image to activate the students’ mental activities.

People use onomatopoeia to describe their facial expressions in daily lives, because onomatopoeia can express the subtle nuances included in facial expressions through a little verbal information. We consider that natural facial expressions of agents or robots can be created by parameterizing facial information included in onomatopoeia. In this paper, we propose a facial expression model for agents using onomatopoeia.

Onomatopoeias have an ability to represent sounds or state of things. Onomatopoeias follow sound symbolism that is a hypothesis that particular sound and a phoneme can give people paticular impressin or image. According to this, similar particular sound and phoneme can make people imagine similar particular image. Urata et al. utilized the sound symbolism and proposed an onomatopoeia thesaurus map, which can visualize semantic relationship among onomatopoeias. The onomatopoeia thesaurus map is constructed by output of a middle layer of a deep autoencoder. Urata et al. reported that onomatopoeias in local area of the map can be semantically similar. But it hasn’t been verified that the map can support sound symbolism. Our study formulated a hypothesis based on Japanese linguistic knowledge about the sound symbolism. The experiment evaluated the hypothesis of the map. The most of the experimental result supported the hypothesis, however a part of it didn’t.

In recent years, researches on natural language processing are progressing rapidly, and many companies and research institutes are proposing new machine learning models. As for a popular machine learning model in natural language processing Word2Vec by Google Inc. has been used as a mainstream model. On the other hand, BERT by Google AI Language is attracting attention now. In this study, firstly, a method of emotional polarity estimation using BERT is proposed and applied to simple sentences of newspapers. Next, the estimation results are compared with those by the emotional estimation method using semantic analysis without machine learning which the authors already proposed. Finally, this paper shows that the newly proposed method using BERT is superior to our previous method through discussion of the comparison.

Previously, various studies have proposed a vehicle control method focused on optical flow(OF) that a human uses to perceive the direction of self-movement in order to make passengers not to feel danger while autonomous driving. However, the previous studies cannot control using OF calculated by driving image. In this research, we focus on optical flows on a retina to realize vehicle control with a higher affinity for passengers. This paper proposes a new method of OF calculation using a spherical camera to imitate the light on the retina and to improve the accuracy of OF. Especially, the proposed method used sensor fusion of the image processing with a spherical camera and an internal sensor.

We have been developing a robot system that promotes the improvement of driving at home. When reviewing driving, we believe that it is possible for people to become aware of their own driving by using their own driving records. In this paper, as first step to extract dangerous scenes for the system, we propose a model for estimating the distance between a vehicle and stop lines by using the inclusion-exclusion integral regression model.

We propose a system to highlight stop lines depending on a danger level when driving. First, road areas are extracted from a video recorded by a dashcam using semantic segmentation. Second, straight lines in the road areas are detected by a probabilistic Hough transform. Third, the distance from each stop line to the vehicle is estimated by using the inclusion-exclusion integral regression model. Finally, each stop line is highlighted depending on danger level.

Three-mode fuzzy co-clustering is a technique for extracting co-cluster structures from three-mode co-occurrence information among such three-mode elements as user-food-ingredient. The conventional model was constructed with an intuitive criterion, which often makes it difficult to select an appropriate set of fuzziness parameters. In this paper, a novel three-mode co-clustering model is proposed by introducing a probabilistic concept into the conventional one, where partition fuzziness can be easily tuned under the guideline of the intrinsic fuzziness degree of probabilistic models.

The fuzzy clustering methods for dimension reduction to improve the classification accuracy of high-dimensional data are examined. In this report, we proposed the fuzzy clustering methods that reduces the dimension at the same time as clustering, and verified the characteristics of each technique using artificial data.

The fuzzy c-means for Spherical Data (FCS) algorithm is used for dividing datum into some groups. It is for complete datum that contains no defect. Some strategy for clustering of incomplete data sets has been proposed for vectorial data sets. In this paper, two strategy for clustering of incomplete data sets is applied to FCS methods.

Students read company information and select company to work from many companies during job hunting. Among them the method of consistency companies selection with criteria for judgment is needed. The purpose of this study is to uncover cognitive skills for consistency companies selection. We hypothesized that if students are able to make consistency of companies selection, then the companies they select will be similar in the information they value. This paper evaluate the consistency in companies selection based on that hypothesis

This paper deals with a quantitative evaluation of the color distribution at station streets and these effects on landscape impressions. The authors took videos during walking along the station streets, and these color distributions were quantified using the L*a*b* color system. We also conducted an experiment in which participants watched these videos and then investigate their subjective landscape impressions using the SD method. As a result, we identified that there are two types of impression factors, stimulus and spatiality, which are influenced by the chroma and brightness of the landscape, respectively.

In a production line management of food factories, a worker performs pre-shipment visual inspection on product images. In order to save the labor cost, we have developed a service to identify the meat type (e.g., barbecue, shabu-shabu, mince etc.) of a packed meat image using ResNet, as well as the merchandising label on the package using a combination of YOLO and ResNet. We, however, had a few errors when identifying for packages of different meat. By applying Grad-CAM to the identification model, we have found out that the model focuses on the merchandising label instead of the meat itself. Thus, we fixed our model by removing the label area on the training images, and it can now identify the meat type even on packages of different meat.

To elucidate the information processing in the brain, it is important to analyze the stability of the activity in a neuronal network for a long-term In addition it is also important to clarify the interaction between the neuronal activity and inputs from external environment. For this purpose, we are developing a neurorobot, performing organ-like behavior based on the response patterns generated by applying current stimulation to the living neuronal network.For providing a living core unit of the neurorobot system, we should clarify the basic characteristics of developmental changes of the network activity. We examined the temporal changes in spontaneous activity patterns from the start of culture in the neural network using the extracellular potential multisites-measurement system. As a result, the number of firing rate in entire network increased from the start of culture, while the number of autonomous-activity-detecting-channels decreased after days in vitro (DIV20) and the number of firing sites increased significantly. The result suggests that the mature neuronal network with over DIV20 is suitable for the stable living neuronal core for the neurorobot system.