A chaotic neural network reservoir has been proposed as an effective method for introducing various dynamics in the reservoir neural network. In this paper, we propose a method for implementing the chaotic neural network reservoir with a switched-capacitor (SC) circuit technique. Such a circuit consists of three SC chaotic neurons with discrete elements. We then confirm the functionality of the circuit through a discrimination task in which the network distinguishes sinusoidal and chaotic waveforms.

In this paper, we propose a method of controlling an unstable oscillation observed in an interrupted electric circuit with focus on the state-dependent switching action. We define the Poincaré map and calculate the differential forms by linearly differentiating the Poincaré map with respect to the initial values. The control gain is derived from the differential forms, and the validity of the proposed method is confirmed from mathematical and experimental viewpoints. The proposed method focuses on the switching timing and it advances or delays the switching action. Therefore, it makes it easy to realize a control algorithm in a microcomputer.

A backdoor attack is a well-known security issue facing deep neural networks (DNNs). In a backdoor attack against DNNs for image classification, an adversary creates tampered data containing special marks ("poison data") and injects them into a training dataset. A DNN model that is trained with the tampered training dataset can achieve a high classification accuracy for clean (normal) input data, but the inference on the poisonous input data is misclassified to the adversarial target label. In this work, we propose a countermeasure against the backdoor attack by utilizing knowledge distillation in which the DNN model user distills a backdoored DNN model with clean unlabeled data. The distilled DNN model can be trained with clean knowledge on the backdoored model because the backdoor is not activated by clean data. Experimental results showed that the distilled model achieves high performance equivalent to that of a clean model without a backdoor.

Deep neural networks (DNNs) are vulnerable to welldesigned input samples, known as adversarial examples. In particular, an attack involving the generation of adversarial examples is called a black-box attack when an adversary attacks without any internal knowledge of the target network. In a simple black-box attack, adversarial perturbations are selected on the basis of changes in output probability when the input to the DNN is slightly changed. Output probability quantization has been proposed as a countermeasure against the simple black-box attack.
In this work, we quantitatively evaluate the effectiveness of this protection method by using the image degradation index and propose a new black-box attack that can overcome the output probability quantization. We conducted experiments to generate adversarial examples using the MNIST public dataset. In the conventional method, if the fourth digit after the decimal point of the output probability is truncated, perturbations that can easily be recognized by humans appear in the adversarial example, and the attack ability decreases. With the new attack method, we find that adversarial examples can be generated with a sufficiently small degradation even if the output probability is truncated after the second decimal place. This demonstrates that the output probability quantization countermeasure against the simple black-box attack is not effective.

Haptic displays have been attracting attention for several decades and a broad range of devices have been developed to allow interaction with virtual objects. Among these, cablebased architectures were researched in order to overcome the lack of flexibility and the intrusiveness inherent to rigid haptic displays. Cable robots allow the utilization of several end effectors, which can be proved to be particularly useful for providing haptic rendering on several fingers for the bimanual manipulation of virtual objects. Nonetheless, interference can appear when cables touch each other during manipulation. Also, some configurations may require a high number of cables to make full use of the frame of the device. Here, we propose a novel architecture of a haptic display based on SPIDAR. Our purpose is to enlarge the usable workspace of an existing device configuration while lowering the number of cables and without hindering the quality of the display. A hybrid planar haptic device architecture with n+1 cables that includes a linear module is introduced. Design considerations and the kinematics used are detailed. Then, simple tasks are performed to verify the usability of the device as a haptic display.

The application of sensor networks that realize multiple M2M services for ICN (information-centric networking) is drawing attention. Thus far, a method of improving the efficiency of data collection using clusters has been proposed. In this method, the load of the cluster members is reduced by constructing routes and caching data only with cluster heads. However, it is assumed that there is no limit of caching capacity in the cluster head.
In this paper, we consider the case of limited caching capacity and propose a transmission control method with hierarchical clustering. In the proposed method, an upper layer cluster is formed with multiple cluster heads of lower layer clusters. Also, in the lower layer cluster, transmission power is controlled according to the distance between the cluster head and cluster members. From the results of performance evaluation, the proposed method is found to be superior to the conventional method in terms of power consumption.

Recently, ICN (Information-Centric Networking) has been applied to wireless environments such as mobile ad hoc network s (MANETs). In a MANET, each node has limited power resources. However, in many protocols, a node sends an interest packet by flooding, which consumes much power.
To solve this problem, a multipath content acquisition method has been proposed. In this method, since a requester node selects provider nodes, it does not need flooding for the transmission of an interest. However, relay nodes in one of the path s cache only half of the content and cannot send any content to the requester unless both parts of the content are cached.
In this paper, in order to increase the probability of caching the complete content, we propose a content acquisition method using overhearing. In this method, adjacent nodes of a relay node also receive content data. They cache the content data if another part of the content has been cached. As a result, the number of nodes that cache the complete content increases.

In an inter-vihicle communication (IVC) system using OFDM (Orthogonal Frequency Division Multiplexing) modulation and a CSMA/CA (Carrier Sense Multiple Access/Collision Avoidance) scheme as a MAC (Media Access Control) protocol, the performance is degraded by hidden terminals. IVC using DS-CDMA (Direct Spread-Code Division Multiple Access) has been researched to remedy this problem. However, when this method is applied, there is a concern about the degradation of communication characteristics due to intersymbol interference. This intersymbol interference is divided into two types: (1) different intersymbol interference and (2) same intersymbol interference. In this paper, it is shown that each interference can be reduced by using two proposed methods, improving the communication characteristics.

In this paper, the turbo code system with hybrid PPM-OOK signaling (HPOS) has been advanced by adding a new signal converter to improve the performance in the free space optical (FSO) strong turbulence channel. The proposed system is extended from the standard turbo code to the punctured turbo code. The bit error rate (BER) performance characteristics of the proposed system in the strong turbulence channel are evaluated by computer simulation. It was found that the proposed system with the new signal converter outperforms the conventional HPOS system. The BER performance of the proposed HPOS system is almost the same as that of the binary pulse position modulation (BPPM) punctured turbo code system, but its transmission efficiency of the proposed HPOS system is about 1.3 times higher.

Previously, a virtual sound image that is produced with multiple electro-dynamic loudspeakers (EDLs) was diffused due to the reverberation nature of the room. A parametric array loudspeaker (PAL) can produce a sharper sound image than an EDL. However, it is difficult to produce low-frequency sound with a PAL, and the sound quality of the virtual sound image also deteriorates. Although the sound quality of the virtual sound image can be improved with a subwoofer, the sound image moves to the direction of the subwoofer. In this paper, we propose a high-presence sharp sound image based on sound blending using PALs and EDLs.

Bone-conducted (BC) speech has a significant advantage as a solution for speech communication in an extremely noisy environment because of its stability against surrounding noise. However, the quality and intelligibility of BC speech degrade, making BC speech difficult to restore. To solve this problem, we propose a method for restoring BC speech with a combination of a linear prediction (LP) model using line spectral frequencies (LSFs) and a long short-term memory (LSTM) model. We evaluated the method using three objective measurements: log-spectrum distortion, LP coefficient distance, and a perceptual evaluation of speech quality. The results of all three measurements show that our method is better than the previous method, which used a simple recurrent network. These results also show that the model can yield speech with better quality when the LP gain is estimated more accurately.

In this paper, we propose a new training architecture for speech enhancement based on deep neural networks. In the proposed architecture, the generative model producing the noiseless speech is trained so as to minimize the difference between two statistical distribution parameters of the clean speech and generated speech. From the experimental results, we verify that the proposed method can provide better results than the conventional method.

To discretize a nonlinear differential equation, we have previously proposed a hybrid method constructed as a combination of the additive and multiplicative Euler methods. In this study, we formulate the vector field for which the hybrid Euler method is effective. Then, we evaluate the method through numerical and physical experiments for a tomographic dynamical system using, respectively, a sinogram synthesized by a digital phantom and a measured projection acquired from an X-ray computed tomography scanner. We found that the hybrid Euler method has an advantage over both the additive and multiplicative Euler methods.

Demosaicking is an image reconstruction process to recover a full color image from color filter array (CFA) mosaic data. Recently, deep convolutional neural network (CNN)-based demosaicking methods have been explored and have achieved state-of-the-art accuracy. In the deep-CNN-based demosaicking, output pixels are affected by a large spatial region; however, the information involved in demosaicking often only exists locally. In this paper, we propose a channel-wise predictive filter flow (PFF) for demosaicking. Since the PFF is a model that predicts a space-variant linear filter that is transformed to the target image by linearly combining it with the input image, target pixels are reconstructed only from local information. To incorporate the PFF into demosaicking, the proposed network synthesizes the filter flow corresponding to each channel by different networks that are learned independently. Experimental results demonstrate that the proposed method provides better or competitive results compared with several state-of-the-art deep-CNN-based demosaicking algorithms.

Cellular neural networks (CNNs) were developed by Chua and Yang in 1988. The CNN concept was inspired from the architecture of cellular automata and neural networks. The performance of a CNN depends on template parameters. CNNs can be applied to various types of image processing by changing the template. However, little has been reported on the sharpening of blurred images by using a CNN. In this study, we propose an algorithm to sharpen blurred images. Then, we apply the proposed algorithm to the input images and investigate the performance by simulations.

This paper describes a statistical correction model for wind speed data of the Meso-Scale Model Grid Point Value (MSM-GPV), which is one of the numerical weather forecasting systems. In the numerical forecasting system, there are calculation errors caused by both the physical modeling and estimation of initial values. Because numerical forecast data have two-dimensional spatial information, convolution with a convolutional neural network (CNN) is used to grasp and correct the two-dimensional features of errors contained in the forecast data. In the simulations, several MSM-GPV data used for the input data and various correction models are prepared and compared with the results of a fully connected neural network from the viewpoints of the error improvement rate and error distribution.

This paper describes a safe driving support system that reduces drift in the downhill direction when an electric wheelchair crosses a slope. To solve this drift problem, we propose a safe driving support system that controls a wheelchair based on the slip condition of drift, which is calculated from the roll angle, yaw angle, and wheelchair speed. Experimental results obtained using the developed electric wheelchair verify the usefulness of the proposed system.

The goal of this research is to detect abnormal behavior in vibration data of factory machinery and take measures against failure. In this study, an accelerometer is mounted on a mechanical device with a conveyor belt. Vibration data are collected to monitor the conveyor belt line. We propose a method of detecting vibration abnormalities by combining signal processing and an autoencoder (AE), one of the neural network models. In the learning phase, the vibration signal is converted to a spectrogram and used for the output of the neural network. Then, a random mask is applied to the horizontal direction of the spectrogram. This technique does not require a search for a valid frequency band. The masked spectrogram is used as the input for the neural network. A model that converts the masked spectrogram back to the original spectrogram is trained. This model is a type of AE, a deep convolutional encoder-decoder architecture. In the detection stage, the masked spectrogram is input to the model and a predicted image is obtained. The predicted images are evaluated with the weighted moving variance of the PSNR. In an experiment, a normal AE and two masked AEs were compared. Both the AE and the proposed method can detect faults, and it was shown that other faults can be detected by a masked AE.