Neural Architecture Search (NAS), which aims to automatically optimize the structure of a neural network for achieving excellent classification performance, has attracted considerable attention in recent years. Recently, zero-shot evaluation methods have been proposed for estimating classification performance without training to reduce the search time. However, these indices are still insufficient for finding the best-performing neural networks. In this study, we demonstrate that it is possible to evaluate convolutional neural networks (CNNs) using the robustness of the rectified linear unit (ReLU) output distribution to weights. We propose a new zero-shot CNN evaluation index based on this robustness index.

In this paper, a burst-based multilayered cortical learning algorithm (BM-CLA) for forecasting trend-changing time-series data is proposed. CLA predicts time-series data while adjusting synapse relationships online. However, the forecast accuracy of the conventional CLA deteriorates with trend-changing time-series data, in which several time-series trends are switched over time. The proposed BM-CLA detects trend changes based on multilayered CLA predictors. Experimental results using multiple artificial time-series data with probabilistically changing trends showed that BM-CLA achieves results that are better or comparable to those of conventional CLAs with different specifications and the long short-term memory (LSTM), which is a neural network-based forecast algorithm.

We investigated a (supervised) deep learning model for the automatic segmentation of manufacturing video data into a sequence of assembly operations. In the action segmentation for general human behavior, various temporal convolutional network (TCN)-based methods have been proposed and demonstrated to have stable performance using temporal features from captured extensive fields of image frames. However, they often make it difficult to detect unusual actions occurring in short durations, such as skipping assembly operations. In this paper, we address this drawback in the existing TCN model by introducing two techniques of differentiable group normalization and jumping knowledge to relax the oversmoothing effect. Then, we empirically show the performance of the proposed TCN variations using a benchmark for operational assembly work segmentation. We also present a preliminary result on irregular work detection.

We propose a novel online algorithm for moving blind source extraction (BSE). The original algorithm is based on the recently proposed constant separating vector mixing model with batch auxiliary-function-based independent vector extraction (CSV-AuxIVE). The CSV-AuxIVE is not suitable for some devices that require real-time processing. In this case, we propose the online-CSV-AuxIVE, which only needs to know part of the mixed signal to process the observed signal sequentially. Then, we verified the effectiveness of the proposed method on a specific task of extracting moving two-speaker signals. In the experiment, the online-CSV-AuxIVE is compared with the online-AuxIVA. The result shows that under the same conditions of the source of interest (SOI), the average source-to-distortion ratio (SDR) of online-CSV-AuxIVE is approximately 1.5 dB higher than that of online-AuxIVA regardless of the changes in the range and speed of the interference (IR).

We propose a deep complex-valued neural network-based beamforming framework for multichannel target speech separation. The deep complex-valued neural network predicts steering vectors and complex ratio masks for speaker signals. The masked signals are then used to calculate the spatial covariance matrices needed for minimum variance distortionless response (MVDR) beamforming. We propose triple-path modeling for mask estimation, which takes both intrachannel and interchannel features into consideration. Our experimental results revealed that the proposed framework achieves better target speech separation performance than do the baseline methods.

In many offices, starting and finishing times of workers are recorded as part of attendance control, but they do not monitor the employees' behaviors during the workday whether they are taking appropriate breaks. In this paper, we propose a behavior recognition system using the object detection algorithm YOLOv5 and the object tracking algorithm DeepSORT and two surveillance cameras. In addition, we propose a person matching method between the images from two surveillance cameras using projective transformation. We will show that the recognition works substantially well by integrating the recognition results by using two cameras.

In the coming Internet of Things (IoT) era, it is important to reduce the volume of data being output by sensors as well as their power consumption. Since conventional image sensor output data for photography are often redundant in AI applications, image sensors that can output lightweight data for use in AI are needed. In this work, we propose a complementary metal oxide semiconductor (CMOS) image sensor (CIS) pixel circuit that can extract intensity gradients without the use of analog memories. The gradients are available for histogram of oriented gradients (HoG) features that can reduce the amount of data used in image classification. We performed experiments to evaluate whether the HoG features calculated by the outputs from our image sensor pixels were suitable for image classification tasks. A support vector machine (SVM) classifier was trained with simulated sensor outputs to evaluate human detection accuracy. We also evaluated the accuracy when the sensor outputs were quantized using "low bit decimation" and "value clipping" to reduce the amount of data. Our experimental results indicated that the highest accuracy of 99.55% was achieved using the 2-bit-width quantized gradients by value clipping and HoG features calculated with 4 × 4 cells.

To devise decontamination plans for soil, we have been developing a portable monitoring system that estimates the distribution of radioactive soil contaminants in the depth direction. The system consists of radiation sensors and an estimator using a convolutional neural network (CNN). The rod-shaped measuring instrument, which can be embedded in the soil, has 20 radiation sensors arranged at intervals of 2.5 cm. In our previous study, to create a CNN that can estimate the depth distribution of radioisotopes (RIs) when RIs are only in one layer, we made a large number of simulation datasets to train and validate the CNN. The trained CNN was able to estimate the depth distribution of RIs from the simulation data. However, the CNN should be improved so that it works well for practical situations such as when RIs exist in more than one layer. In this paper, we present an improvement of the CNN so that it can work for RIs in three layers at maximum.

The use of the intelligent reflecting surface (IRS) is being considered for wireless communications in the high-frequency band. For the prediction of IRS reflection patterns, a method that accounts for receiver movement by treating the partially observed channel estimates as a time series from the current time to a certain period of time in the past has been studied. However, the method does not take into account the movement speed of the receiver, which degrades the prediction performance when a variety of speeds exist. Therefore, in this study, a prediction method that uses channel amplitude differences as an indicator to detect the movement speed is proposed. Numerical experiments show the improvement in the prediction performance of the proposed method.

In the 5th-generation mobile communication system (5G), a multiuser multiplexing scheme, nonorthogonal multiple access (NOMA), has been proposed to accommodate the massive connectivity in wireless communication systems. By nonorthogonality, NOMA allows the use of a new domain that is ineffectively utilized in conventional orthogonal multiple access (OMA) schemes. In power-domain NOMA (PD-NOMA), user signals are allocated different power levels and multiplexed by superposition coding at a transmitter. NOMA employs interference cancellation schemes such as successive interference cancellation (SIC) at receivers to avoid interference between the signals. The use of nonlinear systems such as low-resolution analog-to-digital converters (ADCs) is favorable in terms of energy consumption since the load of SIC processes increases proportionally to the number of users. However, on the other hand, the use of nonlinear systems leads to the degradation of signal detection performance. Stochastic resonance (SR) is a nonlinear phenomenon that improves the response of nonlinear systems to weak power signals. In this paper, we propose the application of SR to the nonlinear system in SIC. Through computer simulations, we show that the proposed method can improve the signal detection performance without losing the advantage of low energy consumption.

The design of Media Access Control (MAC) protocols for the IoT has become important owing to the rapid growth in the number of IoT devices. To realize efficient and large-scale wireless networks, autonomous distributed time division multiple access (TDMA) has been studied in previous research. We previously proposed a distributed TDMA suitable for the massive IoT network, which is based on the desynchronization phenomenon of a nonlinear oscillator network with intermittent and inhibitory couplings. In this paper, we propose a performance improvement method for our desynchronization method by changing the coupling functions of nonlinear oscillator networks. Our simulation results demonstrate that the convergence time to achieve a desynchronization state for TDMA communication can be reduced by modifying the coupling functions, such as the linear and cosinusoidal functions.

Recently, sensor devices have been spreading rapidly owing to the development of the internet of things (IoT). They are often battery-driven and therefore have limited energy capacity. To extend the operating time while maintaining the communication function, simultaneous lightwave information and power transfer (SLIPT) is in demand. The aim of this paper is to evaluate the bit success rate (BSR) and power transfer rate (PTR) of Variable xASK-CodeSK by using theoretical formulas. The results indicate that there is a tradeoff relationship between BSR and PTR. Moreover, it is found that there is an optimal value of x for optimizing the tradeoff relationship.

We describe a wind speed prediction method using wind vector images as input. The prediction model combines convolutional neural network (CNN) and convolutional long short-term memory (CLSTM), which are effective for image analysis. Several input image data structures expressing wind vector change are considered and the prediction accuracy is compared between them. The performance of the proposed method is evaluated by the root-mean-square error and correlation coefficient between observed and predicted values.

We propose attitude stabilization control for a biped hopping robot with springs and a closed-link leg mechanism. The robot has an inertial rotor to stabilize the attitude in combination with the swinging control of the leg. The attitude stabilization control is realized using an integral optimum servo controller designed on the basis of the linear quadratic regulator method. The attitude of the robot is estimated from the acceleration of the robot body measured by an IMU. Experiments using this system are conducted to confirm the attitude stabilization performance compared with the simulation results.