One of the problems when using the dialog system is that users do not know what utterances the system can understand to convey their preferences. To mitigate the problem, we developed an associative dialog system for recipe recommendation which utilizes food knowledge base. In addition to providing expressions candidates that system can accept, it also repetitively generates and respond with linguistic expressions related to foods associated with user input.

In this research, I conducted "a persuasive game" where a worker convinces the character about the importance of the work and analyzed the increase in motivation levels of the worker. When the worker fails to persuade, the worker confirms the importance of the work, so the motivation for work will be improved. In this experiment, I prepared two types of systems : one with a character that accept any persuation(conventional system) and one that may or may not accept a persuation , incorporating the success or failure of a persuation based on its cntent (proposed system), compare and evaluate each went. As a result, it was shown that the system that incorporates the success or failure of persuasion tends to significantly improve the motivation before work, compared with the conventional system.

This paper presents a method for acquiring utterance candidates for chat-oriented dialogue systems in specific domains from Wikipedia articles. A previously proposed method extracts sentences from Wikipedia articles related to the domain, converts them to utterances suitable for chat, and excludes sentences inappropriate as system utterances using a classifier. The classifier uses position information of the sentence in the article, but it is based on a simple logistic regression. On the contrary, our proposed method uses a state space model in a Bayesian approach. Experimental results show the proposed method outperforms the previous method.

We propose a dialogue system for children with emotions. In this paper, we focus on utterance understanding and emotion generation in this dialogue system. In utterance understanding, we extract utterance intention and focus. Machine learning based on utterances collected from college students gives an accuracy of about 70% for utterance intention analysis and about 90% for focus extraction for children's utterances. In emotion generation, we deal with two types of emotion: short emotion and mood. The type of response is affected by mood, and short emotion is expressed as a facial expression. Regarding expressions of emotion, college students judged about 60% of emotions generated in dialogue as valid, but we could not confirm their usefulness in dialogue with children.

In recent years, there are many dialogue systems selecting an appropriate response from a dialogue database. However, when the number of sentences in the database is small, it may not contain an appropriate sentence in an open domain setting. Therefore, it is desirable to prepare a large number of candidate sentences and to fully expand the dialogue database beforehand. In this research, we generate templates from handwritten responses from actual dialogue logs to augment interrogative sentences for a dialogue database.

In a human-human conversation, a listener frequently conveys various intentions explicitly/implicitly to the speaker with reflexive short responses. The speaker recognizes the intentions of these feedbacks and changes the utterance plans to make the communication more smooth and efficient. These functions are expected to be useful for human-system conversation also, but when human face the system they do not always give the same feedback as they did with human. We investigated the feedback phenomena of human users against our system which is designed to transfer a massive amount of information like news articles by spoken dialogue. First, user's intentions of feedbacks that can affect the behavior of the system were classified from the viewpoint of roles in the progress of conversation: presence/absence of a user's interest, comprehension state, release/keep turn, and so on. Then, using our news delivery conversation system, we gathered user's short responses, and labeled the intention through listening test by the subjects. We also attempted automatic identification of those intentions. As a result, in the current system, it is found that there are many feedbacks to convey doubt and automatic identification of them is possible stably, but there are not so many other feedbacks.

The purpose of this paper is to propose Jupiter, a new environment for automated negotiations in which we can easily create agents that is able to use machine learning. Genius is cited as a prior study of the environment where automated negotiation can be simulated. It provides an environment for automated negotiations that aim to solve multi-issue negotiation problems. In the field of automated negotiation, it is expected that agreement results are optimized by machine learning. However, it is difficult to use Genius to simulate automated negotiations with agents using machine learning, because the past negotiations information provided by Genius is insufficient. As above, we propose Jupiter as a new automated negotiation environment in which we can easily create agents that is able to use machine learning. In addition, we compare Jupiter with Genius and show the superiority of Jupiter.

Towards creating an argumentative dialogue system, this paper proposes a method to create premises from a given utterance representing a conclusion. The proposed method utilizes Walton's argumentation schemes for generating premises from various inputs. The generated premises were evaluated by human subjects, and the results showed the validity of our method.

We propose a method to create metaphors using corpus and thesaurus. Our system can create metaphors based on input tenors and grounds. We use word2vec to obtain word vectors for tenors and grounds, which are used to select vehicles considering cosine similarity. In order to output various vehicles that are congenial to the ground, we create a grounds-by-categories matrix using thesaurus. We evaluate appropriateness and usefulness of vehicles that are created by our system. As the result, it is shown that our system can output metaphors that are useful for user, though many of the metaphors created by the system are inappropriate. We compared vehicles created using Aozora Bunko corpus and NWJC. As the result, it is shown that our system can output vehicles that are congenial to user's literary style by using appropriate corpus.

Although we have been studying diverse elements of narrative generation, many of the elements have not been combined with our basic narrative generation system called Integrated Narrative Generation System (INGS). In this paper, we list these elements and especially focus on the part of the attribute information of noun concepts in the noun conceptual dictionary in the INGS. We introduce the attribute information into the noun conceptual dictionary to aim at various effects including detailed conceptual explanation generation and rich narrative character formation.

Growing number of robots and increase in its capabilities impose new demands on meta robotic systems, capable of performing hardware abstraction and sharing skills between different robots. In this work such systems are studied, in order to make evident what can actually be done now and what are the problems that must be dealt with in the future.

In this paper, for stable learning and faster convergence in Reinforcement learning continuous action tasks, we propose an alternative way of updating the actor (policy) in Deep Deterministic Policy Gradient (DDPG) algorithm. In our proposed Hybrid-DDPG (shortly H-DDPG), at one time step actor is updated similar to DDPG and another time step, policy parameters are moved based on TD-error of critic. Once among 5 trial runs on RoboschoolInvertedPendulumSwingup-v1 environment, reward obtained at the early stage of training in H-DDPG is higher than DDPG. In Hybrid update, the policy gradients are weighted by TD-error. This results in 1) higher reward than DDPG 2) pushes the policy parameters to move in a direction such that the actions with higher reward likely to occur more than the other. This implies if the policy explores at early stages good rewards, the policy may converge quickly otherwise vice versa. However, among the remaining trial runs, H-DDPG performed same as DDPG.

In develping intelligent robot systems, it is necessary for designers to make robots by fine-tuning parameters and modifying programs in detail. However, this conventional method is a big burden not only for designers but also general users. In this paper, we present a new design method based on reinforcement learning in real time through human-robot interaction. We also develop an autonomous mobile robot system and show its possibilities.

We propose a person re-identification method for mobile robots that periodically provides services to specific groups. This method consists of a feature extractor that learns to extract person features based on the Triplet Loss from person regions detected by a region-based CNN and a person re-identifier that learns to identify persons through transfer learning of person features while moving around a room. The person re-identification incorporates adaptive transfer learning to periodically re-learn the same persons with different appearance, such as clothes etc. Performance of the proposed method is evaluated by an experiment using a public large-scale data set and an experiment using the self-made dataset periodically collected for the same group by the mobile robot.

Making programs play games has contributed significantly to the advancement of AI research. It is partly because popular games often resemble real-world problems. Tackling them has enabled AI to cope with real-world scenarios as well. Now that AI has surpassed humans in strategy board games such as chess and go, one next target would be to train it to play video games. This paper focuses on a shooting game and optimizes a program to avoid bombardments deployed by the opposing player. Using a genetic algorithm, the AI player was optimized to move around without hitting enemy attacks. The results of experiments showed that it can successfully learn to do so, although with much computation time.

Obtaining a human-level control through reinforcement learning (RL) requires massive training. Furthermore, a deep learning-based RL method such as deep Q network (DQN) is difficult to obtain a stable control. In this paper, we propose a novel deep reinforcement learning method to learn stable controls efficiently. Our approach leverages the technique of experience replay and a replay buffer architecture. We manually create a desirable transition sequence and store the transition in the replay buffer at the beginning of training. This hand-crafted transition sequence enables us to avoid random action selections and optimum local policy. Experimental results on a lane-changing task of autonomous driving show that the proposed method can efficiently acquire a stable control.

In this paper we propose a novel method which incorporates self-training into a sequence-to-sequence model in order to improve the accuracy of the headline generation task. Our model is based on neural network-based sequence-to-sequence learning with an attention mechanism and trained with approximately 100,000 labeled examples and 2,000,000 unlabeled examples. Through experiments, we show our proposal significantly improves the accuracy and works effectively.

ATM (Automated Teller Machine)-related crimes have been reported as serious problems overseas. One of such crimes is skimming: the act of installing skimming device on ATM to illegally copy information of cash cards or passwords. Detecting such illegal acts promptly is important since we can immediately prevent ATM-related crimes by automatically raising alarms and also reduce supervisor’s monitoring burden. We have realized an anomaly behavior detection technique which finds not only suspicious devices but also detects actions of installing them. In this article, we discuss our proposed method which combines video feature extraction, ATM status acquisition and recognition. Compared to processing only image information, the proposed method proves to be more effective in recognizing anomaly behaviors from normal behaviors.

There is a movement to try to utilize the drone for checking inside the tank / hold of the ship. Access to the inspection site becomes easy, while on the other hand, since the site is judged through images taken by a drone, a technique for supporting state evaluation on the screen is desired. In this research, we applied the detection technology using deep learning (Faster R - CNN) to the recognition of structural members in hold of bulk carrier as the first stage. Because we have few experience of flying drone in tank / hold, here we learned and recognized based on the images taken at inspection. In addition to recognition in a bright environment, we also conducted virtual experiments simulating the dark environment inside the tank, verified and interpreted the recognition rate, and arranged issues for practical application.

OpenPose of CMU can realize simultaneous posture recognition of multiple people in real time by simple Web camera. We developed two application systems utilizing OpenPose. The first one is a basketball free throw prediction model with human posture data recognized by OpenPose. The second one is a healthcare application which aims to detect brain stroke tendency. In this article, we focus on what we achieved through our first application development. We accomplished to build a high quality motion prediction model with OpenPose data. We consider it opens a way for other research and application developments which are related to human posture and motion.

Japanese sometimes leave the decision of things to the atmosphere, it is very important to design the atmosphere of the place in various situations. BGM has an emotion inducing effect and an image inducing effect, which makes it possible to change the atmosphere of the space. In this paper, we proposed a BGM recommendation system reflecting the atmosphere in the shop by combining images and environmental sounds and constructing a data set using shop labels. As a result of recommendation by our system, higher ratings than baseline were obtained for three index.

The impression of songs varies from person to person. Sensitivity depending on this individual is greatly different between individuals and quantification is difficult. In addition, since sensitivity changes with time, fixed point observation is also necessary and takes time. In this case, focusing on the graph Laplacian which has the property of Fourier transform among the deep learning attracting attention in recent years, we clarify the trend of the frequency characteristic which affects the experiment time by computer experiments and the sensitivity of music.

In recent years, in the field of manufacturing industry, there is an increasing need for anomaly detection by collection and analysis of sensor data for predictive maintenance of factory equipment and industrial product.In the conventional method, for example, attention is focused on the power of a specific frequency band related to the anomaly of the equipment, and anomaly detection is performed by comparing the magnitude of the power.However, when there are many or wide frequency bands involved in anomaly, they may not be clearly specified. Therefore, we propose a method for anomaly detection using non-negative matrix factorization which is an unsupervised learning algorithm. Even if the frequency band of the vibration data is not clearly specified, our proposed method can automatically extract the features of vibration data and detect anomaly of machines accurately.

This paper proposes an anomaly detection system for time series data by using 1d CNN-LSTM networks, which is a combination of one dimensional convolutional neural network (1d CNN) and long short time memory (LSTM). The CNN processes a time series data from a sensor and its output is passed to LSTM. After training of the networks, the CNN grows into an appropriate filter to emphasize features of the time series in frequency domain, and the LSTM becomes a good feature extractor. The extractor works fine even under noisy conditions. It generates a feature vector and the vector is utilized to diagnose anomalies. We applied the proposed system to vibration sensor data obtained from a control valve. The system detected anomalies due to cavitation, which is one of the most serious phenomenon of control valves, with 99.5% accuracy. The result shows availability of our proposed method.

We describe the construction and practice of collaborative learning environment to develop the dementia care skills focusing on elderlies personality. In our learning environment, the policy of care for people with dementia is considered based on their personality and desire. And our learning environment helps to learn collaboratively dementia care throughout the organization by using video. We have conducted a collaborative learning for dementia care in the care fields. The results suggest that the collaborative learning environment is useful for learners.

Many studies reported eye-related movements, e.g., eye blink and eyelid drooping, are highly indicative symptoms of drowsiness. However, few has investigated the computational efficacy for drowsiness estimation accounted by these movements. This paper thus analyzes two typical movements: eyelid movements and eyeball movements, and investigates different neural-network modelings: CNN-Net and CNN-LSTM-Net. Experimental results show that using joint movements can achieve better performances than eyelid movements for short time drowsiness estimation while using eyeball movements alone perform worse even than the baseline (PERCLOS method). In addition, the CNN-Net is more effective for accurate drowsiness level estimation than the CNN-LSTM-Net.

Expense on elderly care is increasing more rapidly than other expense on social security benefits. To Operate long-term care insurance system, improving operating efficiency is the critical issue. Especially, we focused on assessment of nursing care necessity degree that is the first step of long-term care insurance system. Assessment productivity will improve by assessing nursing care necessity degree periodically, objectively and automatically as possible. In this paper, we created nursing care necessity degree classify model using by assessment data for planning care-plan, and analyzed mis-classify trends. We found assessment data for planning care-plan are also useful information on assessing nursing care necessity degree, and there is possibility of improving assessment productivity.

This paper describes an attempt to visualize clinical data collected from multi-institutional EHR System.

In the self-repair networks, a ring network has been studied. However, the network structure in the real world is more complex and diverse. This research discussed the boundary between the spread phase and the eradication phase of abnormal nodes by Mean Field Approximation (MFA) in the regular graph of degree three. We also conduct agent simulations for the regular graphs with degrees exceeding three. As a result, we have derived parameter region that guarantees abnormal nodes extinction.

Materials informatics (MI) is a promising approach to liberate us from the time-consuming trial and error process for material discoveries. Contrary to molecular systems, however, practical successes of the solid-state MI are very scarce because existing descriptors insufficiently describe 3D features of eld quantities (e.g., electron distributions and local potentials). We develop a simple, generic 3D voxel descriptor that compacts any field quantities, in such a suitable way to implement convolutional neural networks. We examine the reciprocal-lattice 3D voxel space descriptor encoded from the electron distribution by a regression task with 680 oxides data. The present scheme outperforms other descriptors in the prediction of Hartree energies that are signicantly relevant to the long-wavelength distribution of the valence electrons.

To automatically extract the features required to predict the material properties from microstructure, we utilized convolutional newral networks (CNNs). Training of CNNs was carried out by using experimentally obtained scanning electron microscope images and ionic conductivities. After training, the nodes, which placed after global average pooling (GAP), specically activated by passing through the images with high/low ionic conductivities were searched. Then, the corresponding feature maps just before GAP operations were shown. We found that the CNNs focused on the reagion containing large voids and on the area without any crystal defects for the specic features for the material with high ionic conductivities while on the small voids for low ionic conductivities. Such observations agree with knowledge of materials engineering.

Aging infrastructure that was developed during the period of high growth is beginning to become obvious. Visual inspection and sensing are carried out to monitor the state of deterioration in preventive maintenance. It is the standard procedure in Japan that bridges with important structures are inspected once in five years where they visually checks the proximity. A high accuracy and efficiency improvement of the bridge inspection is an urgent issue for the both local public municipalities and private enterprises. This paper proposes multi-grade classification models to visualize the weights of feature variables that aim the countermeasure. In addition, the author proposes a method of sensitivity analysis on deterioration progression in the future in response to unit change of degradation factors. Indeed, these methods applied to a bridge inspection data such as super structure steel corrosion and deck reinforced concrete crack

It is said that we are in a state of mind wandering for approximately 50% of the day. Mind wandering while driving sometimes cause traffic accidents. In order to perform mindful driving, our attention should be appropriately directed toward the objects surrounding the driver, but should not be captured by them. For that reason, it is important to detect mind wandering. In this study, drivers' mind wandering was dened from their behavior during simulated driving and their brain activity was measured and investigated using functional near-infrared spectroscopy (fNIRS). Fractional amplitude low-frequency uctuation(fALFF) was used as an indicator of brain activity. From the results, mind wandering while driving can be detected by variations in the steering angle along with brain activity in the forehead.

In this study, to automatically detect underground objects from the ground penetrating radar (GPR) images by the deep neural network (DNN), we have generated GPR images for training the DNN using a fast finite-difference time-domain (FDTD) simulation with graphics processing units (GPUs). Also we have obtained characteristics of underground objects using the generated GPR images with a convolutional neural network (CNN) and finetuning using a modified VGG16 trained by the ImageNet. It is shown that the CNN and the VGG16 can identify four materials of experimental GPR images roughly 75 % and 80 % accuracy, respectively.

In recent years, the development of autonomous operation technology has been actively carried out in a various research institutions and companies. Many experiments are conducted in public road to confirm whether one autonomous car can drive safely. However, autonomous operation with inter-vehicle communication has not been much researched. In this paper, we implement mutual concessions of autonomous cars with Deep Q-Network (DQN), which is a deep neural network structure used for estimation of Q-value of the Q-learning method. To verify the effectiveness of mutual concessions, we develop an experiment environment for verification of autonomous operation with radio control (RC) cars. We implement mutual concessions of autonomous cars at the confluence in the roundabout. DQN is applied for the decision-making mechanism to decide velocity in the roundabout based on the position of others and the status of congestion. As a result of our experiment, we confirmed that the autonomous cars with DQN can realize high transfer efficiency in the roundabout.

Comics consist of panels, drawn images, speech balloons, text, and so on. In this work, we focused on the difficulty to design text which used for the narration and quotes of characters. In order to support creators to design text, we propose a method to design text by font fusion algorithm with arbitrary existing fonts. In this method, users can change font type freely by indicating a point ion the font map. In addition, we implement the prototype system and discuss the usefulness of our system.

Real world data are often difficult to obtain. Logical machine learning methods can produce perfect explanations for dynamics of systems when the full state transitions can be observed, but such scenario is often impossible. Statistical machine learning methods also usually require a huge amount of data. In this work, we propose a method that predicts the initial weight of an MLP to learn a model that can predict future state of a delayed system even when only a limited amount of observation is provided. We also show the effectiveness of the method applied to systems with particularly a large number of variables.

In recent multimodal learning, deep neural networks are increasingly used as discriminators. In general, we need a large amount of labeled dataset for training them, but it takes a human cost to label multimodal inputs. Therefore, semi-supervised learning on multimodal data becomes important. Among these methods, semi-supervised multimodal learning with deep generative models has recently been proposed. In this study, we first compare these methods and show that SS-HMVAE, which is a method with latent variables corresponding to joint representation, have high performance when different modalities have no deterministic relation in particular. Next, to predict labels from a unimodal data, we propose SS-HMVAE-kl that is an extended model of SS-HMVAE. We confirmed that this method greatly improves the performance when inputting a single modality compared with the conventional models.

Recent expansion of intelligent gadgets, such as smartphones and smartwatches with vital sensors, make it easy to sense a human behavior. We are developing a road accessibility evaluation system inspired by human behavior sensing technologies. Our proposed system aims to estimate road accessibility as environmental factors, e.g. curbs and gaps, which directly influence wheelchair bodies, and human factors, e.g. wheelchair users’ feeling tired and strain, which are results of the environmental factors. This paper introduces a data labeling method using GPS for DCNN learning to extract road surface characteristics from wheelchair sensing data. As a conventional method, the manpower based labeling have been used by comparing wheelchair sensing data with recorded video of wheelchair traveling. This paper evaluates and reports the effectiveness of the proposed method.