Estimation of learners' mental states during the interaction between teacher and learners is very important issues for teacher from quality of learning environment point of view. In this experimental study, relationship between teacher's utterances, behaviors, learner's physiological indexes and mental states were tried to be detected by Machine Learning Method. Particularly, in this study, we tried on estimation of learner's mental states from physiological indexes considering time dilation and persistent model of mental states. As a result, the effectiveness of considering persistent model and time delay of physiological indexes were suggested.

Robots as learning partners could bring about the authenticity of learning contexts, which could also enhance learners’ engagement and experiences. In particular, it is required in the contexts of communication, presentation, and reflection of behavior. This paper demonstrates three learning environments with robots, which are partner robot for collaborative reading in English, presentation robot for lecture, and presentation avatar for self-review, and discusses how these robots could provide the learners with more fruitful learning experiences and more cognitive gains.

This paper investigated whether expressing gratitude and being expressed gratitude via voice agent interaction can improve users' happiness similar to writing and reading gratitude. The experimental results of 22 people in two groups, expressing gratitude and being expressed gratitude via voice agent, showed that both groups improved happiness which was measured by psychological scales. It was a interesting result that about 50% people of the group of expressing gratitude complained of the negative feelings like shame on expressing gratitude, in contrast to the group of being expressed gratitude.

AI research on incomplete information games has aspects that imitate human intelligence. The author chose a card game called Hanabi as an evaluation of imitating human intelligence with AI. Hanabi is a cooperative and incomplete information card game, and this game has a unique feature that a player cannot observe his / her own cards. Player selects whether to build a set of cards or discard the card while providing hints of the card with the cooperator. In the previous research, it was shown that the imitation of human behavior that corrects incomplete information increases the score in Hanabi play between agents. The author evaluated agent's function to modify incomplete information based on behavior of cooperator in the game with human and agent. I experimented with human and two kinds of agents with a difference whether to imitate modify incomplete information as a cooperator of experiment participants and I analyzed the game results and impression evaluation of participants. The result is that while the imitation of reflex intelligence is not effective for increasing the score, the participant has a good impression on the agent when the agent modifies the incomplete information properly, and the difference with the case of human collaborator was not seen.

Data collection is one of the essential tasks of non-task-oriented dialogue systems. Role-playing question answer- ing is proposed as an effective data collection approach to collect a large amount of consistent QA data at low cost. This QA data can use for example-based non-task-oriented dialogue system practically. However, collecting not only role-play QA data but also more diverse examples is important to answer various users ’questions. We propose an extension technique that increases the number of examples for non-task-oriented dialogue systems. This technique helps to erect non-task-oriented dialogue system, which can reply to question that does not include in role-playing QA data. In this research, we make a large number of consistent examples using a small amount of role-playing QA data and a large amount of Twitter data.

This paper examines the applicability of the reinforcement learning schema for modelling player's decision-making process within a signaling game context where one player has information the other player does not have. This situation of asymmetric information is very common in the realworld. Though many applications of signaling games have been developed to solve economic problems, the previously proposed models could not reproduce the human way of signaling. We show some interesting empirical results concerning the refinement of equilibria by the proposed reinforcement learning model.

We introduce a new determination method for the full rankness of a matrix over the binary field, which is made with algorithms used by SAT solver as a reference.

Network design is an important issue for several services and systems such as transportation and telecommunication, and is formulated as network design problem using graph structures. Given a graph and constraints, the goal of the problem is to find a constrained subgraph that satisfies given constraints. For some representative constraints, many algorithms have been proposed. However, they are not enough in some applications, because more general and non-mathematical constraints can be required. Therefore, we aim to obtain a set of all constrained subgraphs for supporting to select a suitable one, and propose a unified approach for network design problems. Our approach describes a set of all constrained subgraphs by an equation using some set family operations. To process described equation, we utilize zero-suppressed binary decision diagrams (ZDDs) that manipulate set families in compact form. We applied our approach to some benchmark instances with representative constraints, and obtained sets of all constrained subgraphs in reasonable time.

Restart strategy is one of heuristics used in CDCL solvers which restarts the search process to make a change to the search space. In this study, we show the empirical results that Glucose restart strategy does not often give good change to the search space after restart. We show that there are frequent patterns of decision variables that will cause bad restarts. This result can give a hint that how to design the variable selection heuristics to give good changes to the search space after restart.

The goal of this research is to construct a conversational android that evokes users' motivation to talk in non-task-oriented dialogue like chatting. It has been said that stating subjective opinions is effective for motivating people to talk; however, users feel it to be unnatural when a conversational android states its subjective opinions. We hypothesized that lacking the background information as to why and how the android has the subjective opinions leads to the sense of unnaturalness because the users cannot accept its subjective opinions without such information. The experimental results showed that stating the background followed by the subjective opinion was signi cantly more natural than the opposite case; whereas, the naturalness was not in uenced by the order if the speaker is a human. These results suggest that sharing background information in advance is an effective strategy for conversational androids to naturally state their subjective opinions.

In order to develop services utilizing robots, various AI element technologies and multiple robots and sensors must be integrated. However, at present, it takes a lot of cost to develop such integrated intelligent applications. Currently, we have been developing PRINTEPS (PRactical INTEligent aPplicationS), which is a user-centric platform to develop integrated intelligent applications only by combining four types of modules such as knowledge-based reasoning, speech dialog, image sensing and motion management. In this paper, we report on a multi-robot teahouse at a university campus festival as a practical application of PRINTEPS. We also report the evaluation of the robot teahouse from the viewpoint of the service quality based on the questionnaires of the customers who experienced the robot teahouse.

In this paper, we propose a method for motion learning aimed the execution of autonomous household chores by domestic service robot in real environments. For autonomous activity by robots in home environment, it needs to define appropriate actions for the environment. However, it is difficult to define theses actions manually. Therefore, body motions that are common to plural actions are defined as primitive motions. Complex actions can then be learned by combining these primitive motions. For learned primitive motions, we propose a reference-point and object dependent Gaussian process hidden semi-Markov model (RPOD-GP-HSMM). For confirmation, a robot perform actions included in several domestic household chores by tele-operation. The robot then learns the associated primitive motions from the robot's physical information and object information.

In the human community, there are various interactions and humans can learn them by observing them or interacting with others. For realizing robots that can coexist with humans, it is important for robots to be able to learn appropriate interactions in the community. In this paper, we propose the novel model coupled Gaussian process hidden semi- Markov model (Coupled GP-HSMM) that enables robots to learn rules of interaction between two persons by observing it in an unsupervised manner. The continuous motions of the persons are segmented into discrete actions based on GP-HSMM, and relationships between the actions are extracted. Moreover, all corresponding actions are not simultaneously conducted by two persons in actual interaction and coupled GP-HSMM models such lags between actions. We conducted experiments using motion data of interaction games and experimental results showed that coupled GP-HSMM can estimate actions, lags between them and their relationships.

Offline handwritten character recognition still remains a tough challenge for AI techniques and algorithms. This is because handwritten documents frequently introduce some amount of noise in the images during the scanning procedures. The presence of noise in the scanned images make them murky and/or blurred and therefore hard to read. In this study, we tried using the CNN architecture named “U-Net” to analyze 607,200 sample images consisting of 3,036 Japanese characters. Our results indicate that the “U-Net” has sufficient ability to get rid of noise from characters and enhance the parts of strokes even though there are a huge variety of handwritten styles.

In the muddy soil pressure balanced shield method, it is necessary to grasp the properties of the muck and to take appropriate measures in order to proceed with stable excavation. Confirming the properties of the muck so far has been done by collecting the muck, performing a slump test, checking with a touching with hand, etc. However, it was dangerous to collect the muck flowing on the belt conveyor. Therefore, the operator sees the image shown on the monitor and discriminates the muck properties. In this study, we tried to determine the muck properties shown in the monitor using image analysis by Deep-Learning such as convolutional neural networks(CNN). Since the possibility of discriminating muck properties can be found by this verification, we will report on the implementation contents and results.

We propose a novel virtual try-on method based on generative adversarial networks (GANs), which we call SwapGAN. Conditional Analogy GAN (CAGAN) has already been proposed as a virtual try-on method based on GANs, though this method is not good at generating with complex patterns of clothing. By considering clothing regions, SwapGAN enables us to reflect the pattern of clothes better than CAGAN. Our method first obtains the clothing region on a person by using a human parsing model trained with a large-scale dataset. Next, using the acquired region, the clothing part is removed from a human image. A desired clothing image is added to the blank area. The network learns how to apply new clothing to the area of people's clothing. Furthermore, an image of the clothes that the p erson is originally wearing becomes unnecessary during testing.

Following great successes of machine learning in various benchmarks, its practical use is attracting attention. The machine learning system has to be trained using a wide variety of data samples and to be tested under various conditions, but collecting numerous data samples is very costly. Here, a demand for data augmentation arises. In this paper, we tackle the augmentation of real images by translating their modality to another modality such as daytime vs. night-time. This data augmentation enables us to train and test the machine learning system in various modality. We first demonstrate that existing approaches, pix2pix and cycle-GAN have some difficulties of applying data augmentation; pix2pix requires paired samples in both modalities or cannot overcome the difference in the modalities, and cycle-GAN sometimes fails in keeping consistency in both modalities. We propose modifications of these methods, which improve the consistency in image modality translation.

Frames are the most fundamental objects in the expression of manga. Many technologies have been developed to identify frame areas from manga graphics, however, there is no universal method that can identify frames as well as human readers can from various types of manga expressions. We propose a hybrid identification method of frames that combines image recognition with microtask crowdsourcing. This hybrid method brings together the accuracy of human readers, and the efficiency of dispersed identification tasks.

We present an sketch generation system composed of Sketch RNN, Google released. It is a recurrent neural network (RNN) constructed stroke-based drawings. Our system is different from traditional way, it is based on serial strokes which human draws. The diference makes output image much smoother and more natural. We introduce Sketch RNN into a tool of sketch generation. This paper focuses on two points. The rst, we check that sketch production with Sketch RNN makes output image similar to input one. The second, we confirm that all the latent vector given by Sketch RNN make a sketch compiled soft strokes. Through these condideration, we explore new sketch generation system.

During a loss assessment after a disaster, a drone is useful to take the photos of the building roof. However, there is a possibility of touching with obstacles and persons when using drone. Therefore, in this study, we developed the system that can control drone in real-time by combining the object detection called YOLOv2 with the drone control program. By outdoor test flights, we confirmed that the system can recognize the person's position and control the drone’s movement according to the position.

We propose a semi-supervised method for estimating the number and locations of products in pictures. Many existing approaches can estimate objects locations in images by supervised learning which needs images annotated with objects locations. On the other hand, our method needs only numbers of objects in images. The experiment shows effectiveness of our method.

For CNN training, abnormal images were drawn using original normal images based on skilled worker’s experience knowledge. CNN could learn features of abnormality from those drawn images. Even in the case of rare abnormality like machine trouble, if abnormal images are visualized with skilled worker’s know-how, CNN will be possible method to detect abnormality without real teacher data.

Natural visual stimuli contain rich “shitsukan”, such as glossiness, translucency, and material of an object. Explaining various types of shitsukan in a unified framework is difficult because in general visual perception involves numerous features. Here we tried to explain visual features for shitsukan perception by analyzing the experimental data of shitsukan discrimination tasks. We assumed that participants responded based on the image features of the stimuli. The features were calculated by a deep neural network (DNN) optimized for image classification, in which more complex and abstract features are represented in the higher layers. The features in the middle layer best explained the participants’ responses, suggesting that relatively complex features are used for shitsukan perception. We also found that the effective features depends on the type of the shitsukan. These results suggest the effectiveness of a DNN for explaining visual features for shitsukan perception.

Many studies on interior have been conducted in various fields, such as simulation of indoor space, optimization, impression evaluation of materials used for walls and floors, and so on. Since the overall ambience and impression of the interior space influences the user's emotions and behaviors, a simple and intuitive interior proposal method according to the ambience requested by the user is required. Therefore, in this study, we focused on the design as a design in the interior, and constructed a system that proposes materials such as walls and floors according to the atmosphere in the indoor space required by the user. At the time of proposal, we input onomatopoeia including fine nuances which can not be specified clearly in other words as words expressing the overall ambience of the interior space, and intuitively proposed the design to the user.

In recent years, deep learning has attracted attention not only as a method on image recognition but also as a technique for image generation and transformation. Above all, a method called Style Transfer [Gatys 16] is drawing much attention which can integrate two photos into one integrated photo regarding their content and style. In this paper, we propose to use words expressing photo styles instead of using style images for neural image style transfer. In our method, we take into account the content of an input image to be stylized to decide a style for style transfer in addition to a given word. We implemented the propose method by modifying the network for Unseen Style Transfer [Yanai 17]. By the experiments, we show that the proposed method has ability to change the style of an input image taking account of both a given word and its contents.

Recently, music studies based on deep learning that require a large amount of input have been garnering attention increasing. Along with that, the task of generating accurate scores from audio data is also important. Although NMF is often used for music factorization into sound basis and activation, there is room for improvement and many methods have currently being proposed. In this paper, we propose method of polyphonic music factorization using RBM. RBM is stochastic model and outputs binary-valued latent features, which is suitable for music score notation. Furthermore, we also propose sparse-RBM in order to settle cross cancel problem. In conclusion, our proposed method showed better accuracy than NMF.