Computation methods using custom circuits are frequently employed to improve the throughput and power efficiency of computing systems. Hardware development, however, can incur significant development costs because designs at the register-transfer level (RTL) with a hardware description language (HDL) are time-consuming. This paper proposes a hardware and software co-design environment, named Mulvery, which is designed for non-professional hardware designer We focus on the similarities between functional reactive programming (FRP) and dataflow in computation. This study provides an idea to design hardware with a dynamic typing language, such as Ruby, using FRP and provides the proof-of-concept of the method. Mulvery, which is a hardware and software co-design tool based on our method, reduces development costs. Mulvery exhibited high performance compared with software processing techniques not equipped with hardware knowledge. According to the experiment, the method allows us to design hardware without degradation of performance. The sample application applied a Laplacian filter to an image with a size of 128×128 and processed a convolution operation within one clock.

Although the classical vector median filter (VMF) has been widely used to suppress the impulse noise in the color image, many thin color curve pixels aligned in arbitrary directions are usually removed out as impulse noise. This serious problem can be solved by the proposed method that can protect the thin curves in arbitrary direction in color image and remove out the impulse noise at the same time. Firstly, samples in the 3x3 filter window are considered to preliminarily detect whether the center pixel is corrupted by impulse noise or not. Then, samples outside a 5x5 filter window are conditionally and partly considered to accurately distinguish the impulse noise and the noise-free pixel. At last, based on the previous outputs, samples on the processed positions in a 3x3 filter window are chosen as the samples of VMF operation to suppress the impulse noise. Extensive experimental results indicate that the proposed algorithm can be used to remove the impulse noise of color image while protecting the thin curves in arbitrary directions.

Fast computation of singular value decomposition (SVD) is of great interest in various machine learning tasks. Recently, SVD methods based on randomized linear algebra have shown significant speedup in this regime. For processing large-scale data, computing systems with accelerators like GPUs have become the mainstream approach. In those systems, access to the input data dominates the overall process time; therefore, it is needed to design an out-of-core algorithm to dispatch the computation into accelerators. This paper proposes an accurate two-pass randomized SVD, named block randomized SVD (BRSVD), designed for matrices with a slow-decay singular spectrum that is often observed in image data. BRSVD fully utilizes the power of modern computing system architectures and efficiently processes large-scale data in a parallel and out-of-core fashion. Our experiments show that BRSVD effectively moves the performance bottleneck from data transfer to computation, so that outperforms existing randomized SVD methods in terms of speed with retaining similar accuracy.

Reinforcement learning methods achieve performance superior to humans in a wide range of complex tasks and uncertain environments. However, high performance is not the sole metric for practical use such as in a game AI or autonomous driving. A highly efficient agent performs greedily and selfishly, and is thus inconvenient for surrounding users, hence a demand for human-like agents. Imitation learning reproduces the behavior of a human expert and builds a human-like agent. However, its performance is limited to the expert's. In this study, we propose a training scheme to construct a human-like and efficient agent via mixing reinforcement and imitation learning for discrete and continuous action space problems. The proposed hybrid agent achieves a higher performance than a strict imitation learning agent and exhibits more human-like behavior, which is measured via a human sensitivity test.

A limited number of types of sound event occur in an acoustic scene and some sound events tend to co-occur in the scene; for example, the sound events “dishes” and “glass jingling” are likely to co-occur in the acoustic scene “cooking.” In this paper, we propose a method of sound event detection using graph Laplacian regularization with sound event co-occurrence taken into account. In the proposed method, the occurrences of sound events are expressed as a graph whose nodes indicate the frequencies of event occurrence and whose edges indicate the sound event co-occurrences. This graph representation is then utilized for the model training of sound event detection, which is optimized under an objective function with a regularization term considering the graph structure of sound event occurrence and co-occurrence. Evaluation experiments using the TUT Sound Events 2016 and 2017 detasets, and the TUT Acoustic Scenes 2016 dataset show that the proposed method improves the performance of sound event detection by 7.9 percentage points compared with the conventional CNN-BiGRU-based detection method in terms of the segment-based F1 score. In particular, the experimental results indicate that the proposed method enables the detection of co-occurring sound events more accurately than the conventional method.

This paper presents a method for many-to-one voice conversion using phonetic posteriorgrams (PPGs) based on an adversarial training of deep neural networks (DNNs). A conventional method for many-to-one VC can learn a mapping function from input acoustic features to target acoustic features through separately trained DNN-based speech recognition and synthesis models. However, 1) the differences among speakers observed in PPGs and 2) an over-smoothing effect of generated acoustic features degrade the converted speech quality. Our method performs a domain-adversarial training of the recognition model for reducing the PPG differences. In addition, it incorporates a generative adversarial network into the training of the synthesis model for alleviating the over-smoothing effect. Unlike the conventional method, ours jointly trains the recognition and synthesis models so that they are optimized for many-to-one VC. Experimental evaluation demonstrates that the proposed method significantly improves the converted speech quality compared with conventional VC methods.

In this paper, we propose a secure computation of sparse coding and its application to Encryption-then-Compression (EtC) systems. The proposed scheme introduces secure sparse coding that allows computation of an Orthogonal Matching Pursuit (OMP) algorithm in an encrypted domain. We prove theoretically that the proposed method estimates exactly the same sparse representations that the OMP algorithm for non-encrypted computation does. This means that there is no degradation of the sparse representation performance. Furthermore, the proposed method can control the sparsity without decoding the encrypted signals. Next, we propose an EtC system based on the secure sparse coding. The proposed secure EtC system can protect the private information of the original image contents while performing image compression. It provides the same rate-distortion performance as that of sparse coding without encryption, as demonstrated on both synthetic data and natural images.

We propose a new framework named ROICF based on reinforcement learning orienting reliable compilation optimization sequence generation. On the foundation of the LLVM standard compilation optimization passes, we can obtain specific effective phase ordering for different programs to improve program reliability.

Cross-project defect prediction (CPDP) is a research hot recently, which utilizes the data form existing source project to construct prediction model and predicts the defect-prone of software instances from target project. However, it is challenging in bridging the distribution difference between different projects. To minimize the data distribution differences between different projects and predict unlabeled target instances, we present a novel approach called selective pseudo-labeling based subspace learning (SPSL). SPSL learns a common subspace by using both labeled source instances and pseudo-labeled target instances. The accuracy of pseudo-labeling is promoted by iterative selective pseudo-labeling strategy. The pseudo-labeled instances from target project are iteratively updated by selecting the instances with high confidence from two pseudo-labeling technologies. Experiments are conducted on AEEEM dataset and the results show that SPSL is effective for CPDP.

In the statistic en-route filtering, each report generation node must collect a certain number of endorsements from its neighboring nodes. However, at some point, a node may fail to collect an insufficient number of endorsements since some of its neighboring nodes may have dead batteries. This letter presents a report generation method that can enhance the generation process of sensing reports under such a situation. Simulation results show the effectiveness of the proposed method.

Synthetic aperture interferometric radiometer (SAIR) is a powerful sensors for high-resolution imaging. However, because of the observation errors and small number of visibility sampling points, the accuracy of reconstructed images is usually low. To overcome this deficiency, a novel super-resolution imaging (SrI) method based on super-resolution reconstruction idea is proposed in this paper. In SrI method, sparse visibility functions are first measured at different observation locations. Then the sparse visibility functions are utilized to simultaneously construct the fusion visibility function and the fusion imaging model. Finally, the high-resolution image is reconstructed by solving the sparse optimization of fusion imaging model. The simulation results demonstrate that the proposed SrI method has higher reconstruction accuracy and can improve the imaging quality of SAIR effectively.

In this paper, we propose a deep model of visual recognition based on hybrid KPCA Network(H-KPCANet), which is based on the combination of one-stage KPCANet and two-stage KPCANet. The proposed model consists of four types of basic components: the input layer, one-stage KPCANet, two-stage KPCANet and the fusion layer. The role of one-stage KPCANet is to calculate the KPCA filters for convolution layer, and two-stage KPCANet is to learn PCA filters in the first stage and KPCA filters in the second stage. After binary quantization mapping and block-wise histogram, the features from two different types of KPCANets are fused in the fusion layer. The final feature of the input image can be achieved by weighted serial combination of the two types of features. The performance of our proposed algorithm is tested on digit recognition and object classification, and the experimental results on visual recognition benchmarks of MNIST and CIFAR-10 validated the performance of the proposed H-KPCANet.

Fresh Tea Shoot Maturity Estimation (FTSME) is the basement of automatic tea picking technique, determines whether the shoot can be picked. Unfortunately, the ambiguous information among single labels and uncontrollable imaging condition lead to a low FTSME accuracy. A novel Fresh Tea Shoot Maturity Estimating method via multispectral imaging and Deep Label Distribution Learning (FTSME-DLDL) is proposed to overcome these issues. The input is 25-band images, and the output is the corresponding tea shoot maturity label distribution. We utilize the multiple VGG-16 and auto-encoding network to obtain the multispectral features, and learn the label distribution by minimizing the Kullback-Leibler divergence using deep convolutional neural networks. The experimental results show that the proposed method has a better performance on FTSME than the state-of-the-art methods.

In this letter, a piecewise linear (PWL) sigmoid function approximation based on the statistical distribution probability of the neurons' values in each layer is proposed to improve the network recognition accuracy with only addition circuit. The sigmoid function is first divided into three fixed regions, and then according to the neurons' values distribution probability, the curve in each region is segmented into sub-regions to reduce the approximation error and improve the recognition accuracy. Experiments performed on Xilinx's FPGA-XC7A200T for MNIST and CIFAR-10 datasets show that the proposed method achieves 97.45% recognition accuracy in DNN, 98.42% in CNN on MNIST and 72.22% on CIFAR-10, up to 0.84%, 0.57% and 2.01% higher than other approximation methods with only addition circuit.

Rehabilitation and evaluation of motor function are important for motor disabled patients. In stride length estimation using an IMU attached to the foot, it is necessary to detect the time of the movement state, in which acceleration should be integrated. In our previous study, acceleration thresholds were used to determine the integration section, so it was necessary to adjust the threshold values for each subject. The purpose of this study was to develop a method for estimating stride length automatically using an artificial neural network (ANN). In this paper, a 4-layer ANN with feature extraction layers trained by autoencoder was tested. In addition, the methods of searching for the local minimum of acceleration or ANN output after detecting the movement state section by ANN were examined. The proposed method estimated the stride length for healthy subjects with error of -1.88 ± 2.36%, which was almost the same as the previous threshold based method (-0.97 ± 2.68%). The correlation coefficients between the estimated stride length and the reference value were 0.981 and 0.976 for the proposed and previous methods, respectively. The error ranges excluding outliers were between -7.03% and 3.23%, between -7.13% and 5.09% for the proposed and previous methods, respectively. The proposed method would be effective because the error range was smaller than the conventional method and no threshold adjustment was required.

The brain processes numerous information related to traffic scenes for appropriate perception, judgment, and operation in vehicle driving. Here, the strategy for perception, judgment, and operation is individually different for each driver, and this difference is thought to be arise from experience of driving. In the present work, we measure and analyze human brain activity (EEG: Electroencephalogram) related to visual perception during vehicle driving to clarify the relationship between experience of driving and brain activity. As a result, more experts generate α activities than beginners, and also confirm that the β activities is reduced than beginners. These results firstly indicate that experience of driving is reflected into the activation pattern of EEG.