In edge computing, edge AI that is oriented to low-latency implementation is attracting attention. Also, with the development of deep learning in recent years, the scale of neural networks implemented on edge AI has been increasing. Therefore, small scale implementation of edge AI is important. On the other hand, individual authentication of semiconductors is urgently needed due to increasing the threat of counterfeit semiconductors. For this reason, an NN PUF has been proposed that implements both Neural Network (NN) and Physical Unclonable Function (PUF) as the individual authentication function of semiconductors. The conventional NN PUF is difficult to reduce the circuit scale due to the large implementation overhead. Therefore, this study proposes a small scale and low latency oriented new NN PUF based on the conventional method. In addition, evaluation experiments using an evaluation board verify the performance for NN and PUF.

The physically unclonable functions (PUFs), which ensure the integrity of devices, have been attracted attention as security techniques for internet of things. The ring oscillator (RO) PUF is one of the most popular PUFs. For security issues of PUFs, the risk of modeling analysis is pointed out. To design the secure authentication system using PUFs, it is important to evaluate the resistance against modeling analysis. Therefore, this study proposes a new modeling analysis using genetic algorithm for RO PUF. Moreover, the proposed method introduces a hybrid modeling method combined a hierarchical method and greedy algorithm in order to improve the modeling accuracy. Experiments showed that the proposed method could realize modeling accuracy of 99% with 50,000 training data sets which was 60% of the conventional method.

This paper studies the effectiveness of the TCP multi-pathization method with SDN for IoT networks on Web services by experiments of QoS/QoE assessment. The experimental results show that the method is affected less by a difference of delay between paths and can always provide higher QoS than normal TCP for Web services. In addition, the authors indicate that the method can provide higher QoE than normal TCP, and suppresses QoE degradation more. From the above results, the paper confirms that the TCP multi-pathization method with SDN is effective for Web services.

This paper studies an appropriate value of the maximum transmission rate under congested for IEEE802.11g wireless LAN by experiments with subjects. In the experiment, the authors studied the appropriate value of maximum transmission rate according to Quality of Experience (QoE) for Web services. This paper treats actual two Web services: Google map and Amazon. This paper studied the QoE-based appropriate values of the maximum transmission rate by experiment. As a result, as Google map and Amazon, we found that the suitable maximum transmission rate exits two point to improve QoE. For Google map, one does not change by the degree of congestion and the other change by it. For Amazon, the suitable maximum transmission does not change regardless of the number of user.

Quantum reading is the protocol that uses entanglement for reading a bit stored in digital memory. By using entanglement, it was shown by S. Pirandola that the data rate was improved. O. Hirota showed that in noiseless environments, error-free reading is possible by using a quasi-Bell state that has maximum entanglement. After that, some non-ideal cases were considered, such as symmetric loss and incomplete phase shift. In this study, we would like to clarify properties of the protocol in more realistic situations for experiment. In this paper, we compare and consider the effects using maximum and non-maximum quasi-Bell states on quantum reading with (1) only non-symmetric loss, (2) non-symmetric loss and incomplete phase shift, and (3) non-symmetric loss and phase noise. In the case (1), we derive the exact solutions of the error probabilities when using maximum and non-maximum quasi-Bell states. Then we clarify that the error probability using maximum quasi-Bell states is always lower than that using non-maximum quasi-Bell states. In the other two cases, we compute error probabilities numerically. In cases (2) and (3), the results are different from that of the case (1). That is, both maximum and non-maximum quasi-Bell states are not always supremer than the other.

An electroporation that formed the pores on cell membrane after application of high electric field pulses has been used the transformation and the sterilization of cell. In this paper, we reported the extraction method of nucleic acid in Escherichia coli and Saccharomyces cerevisiae using high electric field pulse application. The sample was 1.0×10^-3 mol/L NaCl aqueous solution contained E. coli and S. cerevisiae. High voltage pulse is applied to the sample at room temperature. These results were shown as the following: (1) When a pulsed electric field of 4~13 kV/mm is applied to E. coli and S. cerevisiae, the estimated amount of nucleic acid increases with the increases of strength of applied pulsed electric field. (2) In the case of E. coli, the 2-3 kbp band was confirmed at an electric field of 5 kV/mm or more, and the 10 kbp band was confirmed at an electric field of 8 kV/mm or more. (3) In the case of S. cerevisiae, the 2-3 kbp band was confirmed at an electric field of 9 kV/mm or more. the 10 kbp band could not be confirmed at 13 kV/mm or less.

To emulate human emotions in robots, the mathematical representation of emotion is important for each component of affective computing, such as emotion recognition, generation, and expression. In a method that learns DNN from unimodality and represents emotions by vectors of continuous values (Emotional Space), the acquired Emotional Space is likely to depend on the modality, and considering the number of dimensions of Emotional Space is necessary. In this study, we aim at the acquisition of modality independent Emotional Space. We propose a method of acquiring Emotional Space by integrating multimodalities on a DNN and combining the emotion recognition task and unification task of Emotional Space of each modality. Through the experiments with audio-visual data, we confirmed in various dimensions of Emotional Space that there are differences in Emotional Space acquired from unimodality, and the proposed method can acquire a modality independent Emotional Space. We also investigated the compatibility of the recognition and the unification score by changing the number of dimensions of Emotional Space. Finally, we indicate the proposed method can acquire sufficient Emotional Space with a small number of dimensions, such as five or six dimensions, under this paper's experimental conditions.

Humans belong to various social groups. Friendship is one of the factors composing social groups. It is important to understand the characteristics of generation and changing friendship networks in order to understand social groups. Previous studies have proposed models for observing friendship transitions and simulated actual friendship networks. There are friendship generation models based on rational and structural choices. Information Access Scope shows a scope that can know hobbies and preferences about others, and related to rational choice. Thereby, the Information Access Scope affects friendship generation. However, previous studies have not examined which Information Access Scope is suitable for a friendship generation model. In this study, we propose a friendship generation model considering Information Access Scope and simulate multiple patterns of Information Access Scope with reference to previous studies. We compared those simulated friendships with actual friendships. As a result, we were able to find an Information Access Scope suitable for the actual friendship.

This paper proposes a method for identifying and classifying the research data cited in scholarly papers, aiming at automatic generation of metadata stored in data repository. This study focuses on URL citations in the scholarly papers. That is, the targets are to identify the URLs referring to the research data and to classify them into tool and data. The method is realized as a multi-class classification (tool/data/others). The method acquires the distributed representations of the URLs from the context around them, and uses them as the input feature. There exists an advantage in that the meanings of URLs can be given based on their surrounding words. This study adopts an approach of computing the meaning of the entire URL from those of the components of the URL. In order to evaluate the performance of the proposed method, experiments on URL classification were conducted. The scholarly papers included in the proceedings of the international conference were used as experimental data. Experimental results have shown the effectiveness of the proposed method for identifying and classifying URLs referring to research data.

Recently, the number of dementia patients has been increasing due to the aging society. Therefore, it is necessary to develop a system that can automatically judge the degree of dementia progression, not to burden the doctor. The authors developed the puzzle game. We found features during the puzzle game to evaluate dementia progression.

Recently, a lot of studies using Deep Learning techniques have been reported in the field of Digital Histopathology. For instance, there are ideas using deep Convolutional Neural Network (CNN) for disease stage classification and segmentation. These methods are expected to reduce pathologists’ work and realize quantitative analysis. However, at the disease stage classification using CNN, even if we can obtain high classification accuracy, it is difficult for us to understand how CNN decides the disease stage. In this paper, we discussed the relationship between features of cell nuclei shape and the disease stage classification using CNN.

Lifting heavy objects with wrong postures can cause a back pain. Therefore, we propose a new method for analyzing the lifting motion using Kinect in order to point out wrong postures in a lifting motion and to generate guidance sentences for correcting wrong postures. The postures that can discriminate between the correct and wrong motions are those at the starting and ending time points in object lifting. Our proposed method consists of extracting the start and end time points of an object lifting motion and generating guidance depending on the knee and underarm angles. The feature of our method is to be able to extract time points from unstable and transient time series during the start and end of a motion with leg analysis that analyzes the inclusion structure of convex pattern shapes in time series. We also evaluate the precision and recall of the classification between the correct and wrong motions in 20 motion data of 10 people, and analyze the factors of classification failure.

The electrocardiograms (ECGs) are often used as barometers of not only the state of the heart but also the state of health. However, due to their high cost and complicated measurement, they have not been used daily at home. Recently, the development of wearable devices has made it possible to easily measure ECGs, so an analysis algorithm of ECGs that can be used as a preventive medicine have been required. With regard to the automatic analysis of ECGs, while there are many studies that use two-category classification for detecting premature ventricular contraction, few studies deal with multiple classification. In this study, a method of four-category classification was proposed: normal heartbeat, premature supraventricular contraction, premature ventricular contraction, and unspecified class. In the proposed method, a model combining the support vector machine and error-correcting output cording was constructed for 13 types of features obtained from ECG signals. The result of the four-category classification shows that classification accuracy was 99.56±0.26%. The result suggests that the proposed method can be used for early detection of diseases and preventive medicine.

The purpose of this paper is to propose a new simulation model of the hydrogen generation reaction by activated aluminum particles and water. Specifically, it was assumed that the particles were composed of blocks and that the reaction proceeded along the blocks. It is considered that there are numerous nano-cracks form on the surface of the activated aluminum particles and the cracks progress from the surface of the particle to the inside. Therefore, hydrogen generation is thought to be closely related to the progress. First, the crack growth with the hydrogen generation reaction was measured using acoustic emission method. Second, the crack growth was calculated using the proposed model. The initial crack position and crack amount on the proposed model were determined based on the laser microscope image. The crack growth coefficients were determined based on the measured deformation-energy during hydrogen generation. The simulation results were consistent with the change in the integrated value of the normalized deformation-energy. Therefore, it was confirmed that the proposed model is effective to describe the crack growth with the reaction of the hydrogen generation.

The purpose of this study is to improve the accuracy of automatic HTML generation from web page design images. pix2code is the state of art in this field. It is consist of design image learning part by CNN and HTML learning part by LSTM. We propose three improvements of adding a word embedding layer, applying VGG16 fine tuning to CNN, replacing LSTM to Bidirectional LSTM or GRU, and introducing attention mechanism. In the experiment, we employed a conventional data set which was used in pix2code and evaluated by a standard natural language generation metric called BLEU. As the results, the one of proposed models that contained the word embedding layer and the attention mechanism scored the accuracy of 99%. It overcomes the result of state of art scored 88%.

Physicians in the medical field have carried heavy burdens of diagnosis because they need to find various diseases of many patients on the basis of various examinations. Recently, to reduce their burdens, deep learning is enthusiastically applied to medical fields. For example, there have been many applications of deep learning to chest CT and X-ray images. However, there are few studies on deep learning for auscultation. Therefore, we aim to build a lung sound classification system using deep learning. Although a large number of data with annotation are generally required for deep learning, it is difficult to collect a sufficient number of lung sounds data. Therefore, we propose some lung sound classification systems with deep learning for efficiently training neural networks with a small number of data. In detail, 1) Mel-Frequency Cepstral Coefficients are used for feature extraction and 2) some pre-training techniques with three types of neural networks such as a convolutional neural network (CNN), long short term memory (LSTM), and convolutional long short term memory (C-LSTM) are designed to realize efficient learning for a small number of lung sounds data. From the experimental results, it is clarified that the proposed pre-training techniques show effective classification performance, and especially, C-LSTM with pre-training achieves higher accuracy than conventional CNN and LSTM.

In this paper, gait motions with a three-wheeled walker were measured and analyzed, when the rear caster of the walker is pivotable or not. From the results, when the rear caster is pivotable, rotational motions of the walker and the waist, which were different from natural rotation of the waist during normal gait motion, were observed.