Porcine respiratory disease gives enormous loss to farmers. The number of sneezing increases when swine are infected with the disease. However, it is not verified that increasing the number of sneezing was occurred by swine influenza. To verifying them, under controlled infecting condition, sounds and movies were recorded. In this paper, the purpose is developing the support system to efficient assigning labels to acoustic events. First, to extract acoustic events, recorded sounds was applied frequency filter and classified by threshold level based on the signal to noise ratio. After that, both the acoustic event and movie was shown at the same time automatically. As a result, 30000 samples of acoustic events were extracted among 14 days. By using this system, observer assigned sneezing label to 67 samples in 3000 samples. The maximum speed of assigning was 200 samples per hour. Therefore assigning labels was efficient by this support system.

Many conventional freshness (quality) measurement methods are separation analysis, and there are a number of problems such as extremely time-consuming measurement etc. in this analysis. Therefore, in this study, attention was focused on elements and organic matter, and tried to quantify the process of degradation of lettuce. Furthermore, from the surface color and moisture measurement, the relationship between freshness (deterioration) evaluation by appearance quality and objective evaluation is grasped and data set and evaluation method for freshness evaluation leading to machine learning in the future were studied. As a result, it became clear that there is a relationship between surface color and internal quality. It is suggested that freshness of lettuce can be quantified and predicted using only surface color information if accumulating experimental data and constructing a relationship between color change and internal quality using machine learning and depth learning.

With the aim of making an IoT device more environmentally friendly, we attempted to redesign the physical layer of an IoT device architecture for a deep neural network. A method for making a deep neural network device for inference execution with a plastic optical fiber is proposed. The device comprises sum-of-product arithmetic units driven by light from an external light source and non-linear activation function units. These units are concatenated in line and form a layered network without optical amplifiers.

In this paper, we discuss about the evaluation of data augmentation to improve the accuracy for detecting plant diseases. Recently, researches on image-based plant disease detection using deep learning have been conducted. The researches require a huge number of training data, however, it is difficult to obtain so much data. Therefore, the authors focus an application of data augmentation to image-based plant disease detection. In many cases, it is known that data augmentation is effective, however, in some cases performance might be worse. As the condition that the performance of data augmentation deteriorates is not clear, the further researches are required. The authors propose to apply Frechet Inception Distance (FID) to the evaluation of data augmentation. In this study, we investigate the correlation between FID score and performance of data augmentation.

We have practiced mechatronics education using LEGO. Until now, we have prepared individual learning environments where each student can solve tasks by trial and error. However, since individual learning has no information exchange between learners, many students ended up solving the minimum task. Therefore, in this research, we realized a system that can easily integrate devices created by individual learning in cooperative learning by converting devices manufactured by LEGO into IoT. In this learning environment, learners not only can integrate real world equipment, but also can easily control the other equipment depending on the state of the equipment. Learners were able to seamlessly participate in cooperative learning after concentrating on individual learning.

Recently, educational-support robots, which support learning attract many people's attentions. In previous research the robot teaches how to solve some questions only. However, it is difficult for learners to improve their applied skill and inquiring mind, because it cannot prompt learners to deliberate. Thus, this study develops a robot which support based on cognitive apprenticeship. The previous study reported that teaching based on cognitive apprenticeship can prompt learners to deliberate in pedagogy. Therefore, learner who was taught based on cognitive apprenticeship can be improved the applied skills and inquiring mind. In this paper, we investigates learning effect and impression effect of learners for the robot which support based on cognitive apprenticeship.

Japanese Poem Card Game is a game in which a reader reads upper sentence of a poem and players get the card with lower sentence of the poem. One hundred poems from Japanese traditional poem set called “Hyakunin-isshu” are used for the game. A game is played with two players, and a player who got more cards than the other player wins the game. Therefore, players try to get the card as early as possible to win the game. Players need to remember lower sentence of the poem, when a reader begins to read upper sentence of the poem. Since some beginning characters of upper sentences are the same, players need to listen to decision character which decides unique card with lower sentence. The decision character changes during a game, because the number of cards with lower sentences which correspond with the same beginning characters of upper sentences decreases. Since it is difficult for a learner to simulate the change of decision character in mind, we propose and developed a support system which simulates the change of decision character. We verified the effect of the system by an evaluation experiment.

In order to encourage elementary school students unfamiliar with Japanese calligraphy to have an assertive attitude towards acquiring basic calligraphy skills, it is imperative to have an application that provides an environment which allows them to practice calligraphy regardless of the time and place. Due to the fact that this application requires an evaluation system that does not need a calligraphy teacher's assistance, this research proposes a method of automatically evaluating a work of calligraphy. All detected strokes within one sheet will be regarded as an object detection problem, therefore, the entire work's evaluation will depend on the detected strokes. Strokes are a fundamental technique in calligraphy, and are a vital component in creating a work of calligraphy.

In elementary schools, NIE (Newspaper in Education), which uses newspaper as teaching materials in classes, is being implemented. However, teachers’ loads are increasing by adjusting time in general classes, selecting suitable articles and preparing support materials. The purpose of this study is to construct a Web news recommendation system for elementary school teachers. This paper analyzes articles in NIE worksheets based on categories and readability, and examines an estimation method of Web news suitable for NIE in elementary schools.

The training program of high school teachers in Japan lacks the chance to practice teaching skills and the admission of classes. As a new way of training and learning platform, we are running a project to develop a virtual school environment. This paper proposes a design of training system for the student teachers. This system is composed of a virtual classroom which has 30 virtual students. Each virtual student is controlled by an autonomous agent, the agents react to the trainee’s teaching performance in two ways. One is a reflective behavior, the other is emotional behavior which decided by the atmosphere generation model.

Recently, automatic text scoring research using machine learning is progressing in an education field. Most of the short description problems have specific model answers. However, about business education, most of them don't have specific model answers. Therefore, we investigated the possibility of classifying short description problem answers in a business education field, to develop a personalized learning system. If automatic classification is possible, it is possible to provide an appropriate retrospective service in response to learner 's answers, such as changing feedback according to the pattern of answers. As a result, we confirmed that we could classify them with high accuracy if the number of answer characters is about 50 and answers can be clearly classified into less than three patterns. Then we could detect the words that contribute to the classification. Moreover, we found that the way of creating problems is one of the important factors to classify answers.

The effect of Learning by Teaching and learning by being tutored was compared for learning linear equations. Three versions of an online learning environment were created: (1) APLUS for Learning by Teaching, (2) AplusTutor that is a cognitive tutor without problem selection adaptation, and (3) CogTutor+ that is equivalent to the traditional cognitive tutor. A randomized controlled study was conducted in two public schools with 184 7th and 8th grade students. The results showed (i) students’ in the AplusTutor condition finished the quiz quicker than students in the APLUS condition, and (ii) there was a notable difference between Learning by Teaching and learning by being tutored in the “effort” students made.

This paper reports a study on Japanese listening training support system by using Anime. The proposed system divides episodes of Anime into scenes by using silent sections of Anime. The system evaluates a level of a scene by using words and expressions in subtitles of the scene. Users of the proposed system are assigned scenes of Anime which level is suitable for the users. We experimented with the proposed system to verify the efficiency of the system. Participants of the experiment watched scenes given by the proposed system. They took tests prepared from previous JLPT listening tests before and after watching scenes. The scores after watching scenes were higher than those before.

Logical thinking ability is one of the most important abilities in our daily life situations. In this paper, we propose logical thinking training systems to improve logical thinking abilities of high school students and college students by using proving mathematical problems. We show that the proposed systems implemented on Android tablets are effective for improving logical thinking abilities of college students by reporting experimental results and questionnaire.

Importance of interaction among students in university classroom has been increasing because acquiring skills through the interaction can be the basis of future collaboration with diverse people. In this article, the author suggested multiobjective optimization of student group formation for smooth and effective groupwork in university classroom based on this argument. The author applied a hybrid of genetic algorithm and particle swarm optimization (GAPSO) to the multiobjective optimization. Also, the author focused on learning performance and attendance of each student and attepted to find group formation so that the distribution of learning performance and attendance in each student group become as homogeneous among the groups as possible. Comparing with hillclimbing algorithm, GA only, and PSO only in the group formation optimization, the optimization algorithm with GAPSO has the potential to find as many optimum as possible in short time.

A person in the world acquires the situation information that constraints possible range of behavior generation, and then determines proper behavior. We consider ``situated intelligence'' as the generation of behavior adapted to each of surrounding situations. To computationally understand the process mechanism of situated intelligence, our methodology is proposed to model and examine a person as a robotic system. Grounded on making an ecologically valid setting of research, our methodology repeats the following: (1) constructing a probabilistic model of situated intelligence by the parameters of behavior and situation, and (2) practically examining a new hypothesis by practicing the model in a real-world setting.

We study automatic planning of electrical isolation with deep learning, one of applications of artificial intelligence to enhance operation and maintenance. Currently, a skilled engineer plans electrical isolation procedure with hundreds of the circuit diagrams and the related documents, taking man-hours. If this task becomes automatic, it is very efficient. A major issue of the automatic planning is much calculation time of electrical circuit simulator, searching billions of electrical conducting paths. We performed a simplified case study of the electrical isolation. We applied a deep neural network (DNN) for dropping in the calculation time. We trained the responses of the circuit simulator to the DNN, constructing an optimized path search algorithm in the DNN. The calculation time of the DNN was shorter by a factor of 560, compared with that of the electrical circuit simulator. There was no significant difference in accuracy between Multi-Layer Perceptron and Graph Convolutional Network.

Quality control at the construction site is subject to differences in results depending on the skill level of the field supervisor and precision collateral is a problem. In addition, to do precise quality check, it is important to find a suitable "Viewpoints". But it is difficult to efficiently train because it is difficult to extract these "Viewpoints". Also, due to the rapid expansion of the construction market in the current metropolitan area, we can not place employees enough, and we can not provide appropriate guidance by OJT (On-the-Job Training), which is the key to technology transfer. In this study, using an eye-tracking device, we tried to find out skilled field supervisors’ "Viewpoints" or how to "recognize situation", and clarify "judgement criteria" that lead to "action". Then, we show how to inherit the extracted behavior process to supervisor with low skill level.

We propose a model to quantitatively estimate quote spoofing in stock exchange markets without any answer labels, for the sake of more efficient and thorough inspection. In our model, density ratio estimation is used to extract unusual trading activities in unsupervised manner. Using market data at Tokyo Exchange and judges by experts, we validate the model and the result indicates that about 50% of half-day grouped trading histories can be ignored of manual inspection with 80% of frauds in the rest half of the dataset.

This paper introduces a business assisting framework with multiple chatbots. In enterprises, business activities are processed with different employees in multiple departments. Most of existing approaches for business assistance are classi ed into (1)an application of individual activity support and (2)large business suite for assisting business activities processed with different employees in multiple departments. However, to improve business activity more effectively, it is important for a business assisting approach to be applied to business activities processed with not only an employee, but different employees in multiple department, and to be easier to use for each employee and exible for changing business process in each department. In this paper, we propose a new business assisting framework with multiple chatbots. In the framework, each department designs and implements the chatbot which assists its business process as approach (1). Then, these chatbots are connected with each other using shared database for assisting business activities processed with different workers in multiple departments as approach (2). We implemented a prototype of proposed framework applied to service order operation and discuss the effectiveness and issue of our proposal.

We have proposed an Intelligent Lighting System which individually provides illuminance and color temperature required by office workers. The Intelligent Lighting System has been introduced in several offices and its effectiveness is recognized. However, the Intelligent Lighting System can’t realize greatly different illuminance and color temperature required by adjacent office workers because of relation between lighting position and office layout. Therefore, in this research, we propose a new lighting control method which treats illuminance and color temperature satisfactory in the area rather than the specific value by using the index of satisfaction. The verification experiment showed that the proposed method is effective for improving the total satisfaction of all officers. Also, it showed that the proposed method is effective for energy saving lighting control.

Various applications of gas sensors have been envisioned in many fields along with the recent development in information and communication technology (ICT). Gas Identification plays a central role in gas sensor applications including artificial olfaction. In the conventional gas identification protocol, however, a strict gas flow control is required to reproduce comparable sensing signals. To eliminate such a severe constraint and identify gas species with an arbitrary gas injection pattern, here we report an analysis approach based on transfer function, which represents the relationship between inputs and outputs (i.e. a gas input pattern and the resultant sensing signals). In this study, we developed machine learning models which can identify gas species from an arbitrary gas injection pattern. Even though the sample gases were randomly injected, we successfully identified solvent vapors by the transfer functions with the classification accuracy of 0.98±0.03. This study provides a versatile data analysis platform which is independent of gas flow control.

This paper presents a method for learning relative spatial concepts and phoneme sequences which represent spatial concepts and objects from utterances without knowledge of words. First, phoneme sequences recognized by a general speech recognizer are divided into words on the basis of NPYLM. Then, parameters of the relative spatial distributions are estimated from the segmented words and location information by MCMC sampling. In the experiments, the result showed that the parameters were estimated correctly by the proposed method. Moreover, phoneme sequences which represent spatial concepts and objects were learned successfully by the proposed method.

Big data is increasingly used as training data for machine learning. However, large-scale data such as big data is not always appropriate as training data at all times. Particularly when collecting data from Web services such as SNS, unspecified number of people using the service can indirectly tamper with data. In this research, we propose a learning method and verify its effectiveness so as to obtain a learning result close to the case without tampering even in an environment in which part of the training data has been tampered with. In this method, learning is divided into two stages, and the reliability of the training data is estimated using the first stage learning result, thereby assisting the exclusion of tamperd data by human beings.

In recent years, Werewolf has attracted attention in the field of Artificial Intelligence(AI). Werewolf is a kind of communication game and it is classified as incomplete information game. Since Werewolf has unknown information, it is one of good strategy to judge unknown information based on the prediction. Information that player can know is limited, and fake may be mixed in it. In this paper, it is aimed to predict under uncertain information by learning the tendency of players from limited information in Werewolf. In order to consider the time series of Werewolf, LSTM was used. In addition, Werewolf was analyzed using prediction model made by learning.

As the domains of reinforcement learning become more complicated and realistic, standard optimization algorithms may not work well. In this paper we introduce a simple mathematical model called RS (reference satisficing) that implements a satisficing strategy that look for actions with values above the aspiration level. We apply it to K-armed bandit problems. If there are actions with values above the aspiration level, we theoretically show that RS is guaranteed to find these actions. Also, if the aspiration level is set to an ''optimal level'' so that satisficing practically ends up optimizing, we prove that the regret (the expected loss) is upper bounded by a finite value. We confirm these results by simulations, and clarify the effectiveness of RS through comparison with other algorithms.

animals learn not only through individual trial-and-error, but also from other individuals. It is known that vertebrates cleverly utilize learning strategies such as copy-when-uncertain and copy-successful-individuals. These strategies can be applied to social reinforcement learning, although their formalizations are yet to be established. We propose a social reinforcement learning algorithm with a very narrow information sharing. The algorithm exploits RS value function that models the satisficing principle for exploration and exploitation.

Predicting conversion rates (CVRs) in display advertising (e.g., predicting the proportion of users who purchase an item (i.e., a conversion) after its corresponding ad is clicked) is important when measuring the effects of ads shown to users and to understanding the interests of the users. There is generally a time delay (i.e., so-called delayed feedback) between the ad click and conversion. In this paper, we propose a nonparametric delayed feedback model for CVR prediction that represents the distribution of the time delay without assuming a parametric distribution, such as an exponential or Weibull distribution. Because the distribution of the time delay is modeled depending on the content of an ad and the features of a user, various shapes of the distribution can be represented potentially. In an experiment on Criteo dataset, we show that the proposed model outperforms the existing method that assumes an exponential distribution for the time delay in terms of conversion rate prediction.

This paper is the first study aiming at capturing stylistic similarity in an unsupervised manner. We construct a novel style-sensitive word vector predicting the target word for giving nearby and wider contexts under the assumption that the style of all the words in an utterance is consistent. We also introduce an evaluation dataset with human judgments on the stylistic similarity between word pairs. Experimental results illustrate capturing the stylistic similarity significantly.

Network embedding is a method to convert nodes in a network into low dimensional vectors. Most of existing works are designed for single-layer networks, however, real world networks are often represented as multiplex networks. Thus, we propose a novel embedding method for multiplex networks, named MELL, which incorporates layer vectors that capture layer connectivity. The experimental results of link prediction tasks show that MELL outperforms all of the existing methods.

Understanding how black-box classifiers predict is important in many applications, especially in medical diagnosis systems. We propose a effective method to quantify contribution of features to EEG classification using the efficient estimation of Shapley Sampling Value (SSV). EEG data have hierarchical features: an electrode signal, signals in various frequency-bands, and amplitude and phase. If contribution of a feature at a higher level (e.g., a signal of an electrode) is very small, contribution of features at the lower levels of the feature (e.g., signals of frequency-bands of the electrode) should be also small. The method prunes such features at lower levels to reduce computational complexity. We verified the usability of the method in two datasets for EEG classification. The result showed the method could reduce computational complexity of SSV by one third, while maintaining high accuracy of the conventional SSV.

Neural fictitious self-play (NFSP) is a method for solving imperfect information games. While methods developed in recent years such as counterfactual regret minimization or DeepStack require the state transition rules of the games, NFSP works without them. In this paper, we propose to exclude the exploration data from the supervised learning component in NFSP and keep the probability of exploration, in order to explore without breaking the average strategy. We show that this change significantly improves the performance of NFSP in a simplified poker game, Leduc Hold'em, and compare the results for different exploration plobabilities.

As the scope of reinforcement learning broadens, optimization becomes less realistic, and bounded rationality that considers the limitations in agents gets more important. Satisficing, the principal model of bounded rationality, models how people and animals explore and exploit. However, there is no efficient algorithm that represents satisficing can be applied to reinforcement learning in general. We apply our satisficing model, reference satisficing (RS) value function, and the global reference conversion (GRC) technique to the broader reinforcement learning tasks than in previous studies. In the three tasks we deal with in this study, RS and GRC work well, while there are some open problems for general reinforcement learning tasks.

Reinforcement learning is generally performed in the Markov decision processes (MDP). However, there is a possibility that the agent cannot correctly observe the environment due to the perception ability of the sensor. This is called partially observable Markov decision processes (POMDP). In a POMDP environment, an agent may observe the same information at more than one state. We proposed a hybrid learning method using Profit Sharing and genetic algorithm (HPG) for this problem.However, Most of real problems can be represented in an MDP environments. In this paper, we improve HPG to adapt to MDPs environments and report the effectiveness of our method by some experiments with mazes.

Crowd behavior has been subject of study in fields like disaster evacuation, smart town planning and business strategic placing. It is possible to create a model for those scenarios using machine learning techniques and a relatively small training data set to identify behavioral. We implemented a BDI-based agent model that uses such techniques into a large-scale crowd simulator, and apply inverse reinforcement learning to adjust agents' behaviors by examples. The goal of the system is to provide to the agents a realistic behavior model and a method to orient themselves without knowing the scenario's layout, based in learnt patterns around environment features.

In banner advertising, Click Through Rate (CTR) is one of the most important indicators to evaluate one advertisement’s quality. Advertisers create massive number of banner candidates in empirical ways, then proceed to actual tests by delivering advertisement to measure each banner’s effectiveness. This process is expensive and therefore our CTR prediction helps reducing online advertising costs. In this work, we propose a method to classify ‘effective’ and ‘ineffective’ advertising banners based on image processing using state-of-the-art CNN. We first focus only on images then conduct experiments including metadata (product, advertiser, etc) to increase the CTR prediction accuracy and demonstrate which metadata is the most influential. Subsequently, each approach is compared to human performance. In the second part of our work, we detect which parts of the image contribute predominantly to increase the CTR by generating heat maps for each classes. This work leads to a deeper understanding of a banner advertising success and helps making decisions on how to improve it.

This paper describes a motion recognition method to reduce recognition error, which has two-layered structure; motion recognition is affected by context estimation in the first layer, and context estimation is affected by motion recognition in the second layer. We introduce an algorithm to integrate the motion recognition by conventional HMM and motion label production by the topic model in the first layer. We also introduce particle filter to estimate and update the context based on the result of motion recognition in the second layer. A set of particles present a probabilistic distribution of motion topics, and motion recognition and particle update procedures are performed on each particle. In an evaluation experiment, we used a sequential motion which is a sequential connection of 33 motion primitives as a long-term observation target. The results showed that the proposed method reduced recognition errors and tracked motion context by topic probability compared with conventional methods.

We are implementing a system that digitizes paper sticky notes. In order to recognize the positional relationship of sticky notes, it is necessary to take a picture from a position where the sticky note can be viewed from the top. However, the distance between the camera and the sticky notes affects the precision of the extraction of the content of the sticky note.When photographing nearby to accurately read the contents of the sticky notes, information around the sticky notes is lost, making grasping the positional relationship between the sticky notes difficult.There is a trade-off relationship between the extraction precision of the sticky notes and acquisition of the positional relationship between the sticky notes. Therefore, in this paper, we estimate the positional relationship between sticky notes by using a visual inertial odometry. By recognizing the position information at the time of photographing in the virtual space and integrating the extracted information after the photographing in the virtual space by the visual inertial odometry, the relationship between the extraction accuracy and the acquisition of the position information in a trade-off relationship problem.

In recent years, a steady increase in ATM (Automatic Teller Machine)-related crimes has been reported overseas. One of which is ATM skimming; the act of installing skimming devices (a.k.a. skimmers) to ATM to illegally copy information from the magnetic stripes of cash cards, credit cards, etc. As skimmers grow smaller and more sophisticated, detecting such devices with conventional sensors is facing great difficulties. With the purpose of strengthening ATM security, we are developing image sensing technologies that detect anomaly behaviors including ATM skimming acts using video feeds capturing the ATM operational area. Machine learning is employed to represent normal behaviors; the degrees of separation from such representation can be used as an indicator for abnormality level. In this article, we discuss the application of well-known methods (Subspace Representation of [Nanri 2004] and Gaussian Mixture Model of [Yu 2006]) to modelling ATM normal behaviors in order to detect ATM anomaly behaviors. Additionally, we also brief several considerations to realize high anomaly detection accuracy in real practice.

Currently, the practical application of artificial intelligence is dramatically advanced by machine learning using big data. Industrial structure reform and the smart society called Society 5.0 are also expected to be realized. In this paper, real world problems in implementation are discussed. In order to understand internal representation of learning result, Future artificial intelligence project is proceeded. The spiral of two difference tasks, optimization to evaluation function and searching the new evaluation function are introduced as a new challenging issue.