Search

Memristors show excellent ductility and low power consumption. They have broad applications in logic circuits, neuromorphic computing and non-volatile memory. Memristors show great potential in true random number generation (TRNG). To verify memristors' randomness, we made two kinds of memristors, measured the randomness of their delays, and investigated their different randomness mechanisms. We did thousands of experiments and verified the feasibility of making a new TRNG using these two kinds of memristors. We also found a new application of memristors to make non-linear device.

One key challenging issues of facial expression recognition (FER) in video sequences is to extract discriminative spatiotemporal video features from facial expression images in video sequences. In this paper, we propose a new method of FER in video sequences via a hybrid deep learning model. The proposed method first employs two individual deep convolutional neural networks (CNNs), including a spatial CNN processing static facial images and a temporal CN network processing optical flow images, to separately learn high-level spatial and temporal features on the divided video segments. These two CNNs are fine-tuned on target video facial expression datasets from a pre-trained CNN model. Then, the obtained segment-level spatial and temporal features are integrated into a deep fusion network built with a deep belief network (DBN) model. This deep fusion network is used to jointly learn discriminative spatiotemporal features. Finally, an average pooling is performed on the learned DBN segment-level features in a video sequence, to produce a fixed-length global video feature representation. Based on the global video feature representations, a linear support vector machine (SVM) is employed for facial expression classification tasks. The extensive experiments on three public video-based facial expression datasets, i.e., BAUM-1s, RML, and MMI, show the effectiveness of our proposed method, outperforming the state-of-the-arts.

Compact neural networks are inclined to exploit “sparsely-connected”convolutions, such as depthwise convolution and group convolution for employment in mobile applications. Compared with standard “fully-connected”convolutions, these convolutions are more computationally economical. However, “sparsely-connected”convolutions block the inter-group information exchange, which induces severe performance degradation. To address this issue, we present two novel operations named merging and evolution to leverage the inter-group information. Our key idea is encoding the inter-group information with a narrow feature map, and then combining the generated features with the original network for better representation. Taking advantage of the proposed operations, we then introduce the Merging-and-Evolution (ME) module, an architectural unit specifically designed for compact networks. Finally, we propose a family of compact neural networks called MENet based on the ME modules. Extensive experiments on CIFAR, SVHN, ILSVRC 2012, and PASCAL VOC 2007 data sets demonstrate that MENet consistently outperforms other state-of-the-art compact networks under different computational budgets. For instance, under the computational budget of 140 MFLOPs, MENet surpasses ShuffleNet by 1% and MobileNet by 1.95% on ILSVRC 2012 top-1 accuracy, while by 2.9% and 4.1% on PASCAL VOC 2007 mAP, respectively.

Compact neural networks are inclined to exploit “sparsely-connected” convolutions such as depthwise convolution and group convolution for employment in mobile applications. Compared with standard “fully-connected” convolutions, these convolutions are more computationally economical. However, “sparsely-connected” convolutions block the inter-group informa-tion exchange, which induces severe performance degradation. To address this issue, we present two novel operations named merging and evolution to leverage the inter-group information. Our key idea is encoding the inter-group information with a narrow feature map, then combining the generated features with the original network for better representation. Taking advantage of the proposed operations, we then introduce the Merging-and- Evolution (ME) module, an architectural unit specifically designed for compact networks. Finally, we propose a family of compact neural networks called MENet based on ME modules. Extensive experiments on ILSVRC 2012 dataset and PASCAL VOC 2007 dataset demonstrate that MENet consistently outperforms other state-of -the-art compact networks under different computational budgets. For instance, under the computational budget of 140 MFLOPs, MENet surpasses ShuffleNet by 1% and MobileNet by 1.95% on ILSVRC 2012 top-l accuracy, while by 2.3% and 4.1% on PASCAL VOC 2007 mAP, respectively.

Despite the increasing investment in integrated GPUs and next-generation interconnect research, discrete GPUs connected by PCI Express still account for the dominant position of the market, the management of data communication between CPU and GPU continues to evolve. Initially, the programmer controls the data transfer between CPU and GPU explicitly. To simplify programming and enable system-wide atomic memory operations, GPU vendors have developed a programming model that provides a single virtual address space. The page migration engine in this model migrates pages between CPU and GPU on demand automatically. To meet the needs of high-performance workloads, the page size tends to be larger. Limited by low bandwidth and high latency interconnects, larger page migration has longer delay, which may reduce the overlap of computation and transmission and cause serious performance decline. In this paper, we propose partial-page migration that only migrates the requested part of a page to shorten the migration latency and avoid the performance degradation of the whole-page migration when the page becomes larger. Experiments show that partial-page migration is possible to significantly hide the performance overheads of whole-page migration when the page size is 2MB and the PCI Express bandwidth is 16GB/sec, converting an average 72.72× slowdown to a 1.29× speedup when compared with programmers controlled data transmission. Additionally, we examine the impact of page size on TLB miss rate and the performance impact of migration unit size on execution time, enabling designers to make informed decisions.

UE-to-Network relay enables the user equipments (UEs) to provide relay assistance for other UEs located in close proximity. Thus, UEs, especially cell-edge UEs can obtain a better connectivity experience. In this paper, we consider the scenario where UEs can communicate with base stations (BSs) in the direct mode or in the UE-to-Network relay mode. We propose a tractable analytical framework for a UE-to-Network relay enabled uplink cellular network and investigate a flexible mode selection scheme. We leverage stochastic geometry to study the transmission power consumption, relay probability and outage probability. The analysis reveals that UE-to-Network relay is capable of reducing power consumption. Simulation results show the accuracy of our theoretical analysis and demonstrate that the proposed framework can reduce the outage probability. As a result, this paper provides guidelines for mode selection scheme in UE-to-Network relay enabled networks.

Caching popular contents at mobile devices can improve the system performance and relieve the traffic burden for the device-to-device (D2D) enabled mobile ultra dense network (UDN). In this paper, the D2D enabled mobile networks are formulated as a 3-layer architecture involving the link layer, social tie layer and preference layer. Based on the 3-layer architecture, a proactive content sharing scheme including content caching, content diffusion and resource allocation is proposed. For content caching, we propose a joint seed D2D user (DUE) and caching content selection algorithm (JSCSA) with graph theory and caching utility. With contents cached at seed DUEs, a content diffusion approach is proposed based on the social tie and D2D service priority, where seed DUEs can proactively deliver contents to other DUEs with strong social tie when seed DUEs have no requesters to serve. Furthermore, a joint subchannel allocation and power control algorithm (JSAPCA) is proposed to optimize the system throughput. Simulations are carried out to illustrate that the proposed proactive content sharing scheme has superior performances compared with the existing works.

Acceleration for the training process of Deep Neural Networks (DNNs) has been the focus of deep learning field. There were many researches of accelerating deep learning on different platforms. Among them, Intel Xeon Phi Coprocessor is a many-core platform which provides both strong programmability and high performance. But previous work about Intel Many Integrated Core (MIC) focused on parallel computing only in MIC. In this paper, we speed up the training process of DNNs applied for automatic speech recognition with CPU+MIC architecture. In this architecture, the training process of DNNs is executed both on MIC and CPU. We apply several optimization methods for I/O and calculation and set up experiments to approve these methods. Putting all methods together, results show that our optimized algorithm acquires about 20x speedup compared with the original sequential algorithm on CPU which uses one core.

UE-to-Network relay leverages the proximity communication between user equipments (UEs) and allows certain UEs to provide relay assistance for others, which can greatly improve system energy efficiency. In this paper, we consider the scenario where UEs suffering from bad channel condition and low battery level can communicate with the base station directly or via the help of other UEs in heterogeneous networks. The optimal power allocation and connectivity among UEs are studied, which aims at minimizing the system transmission energy while guaranteeing the minimum data rate requirement of each UE. An optimization framework is presented to formulate the system transmission energy minimization problem, which can be converted into a weighted one-to-one matching problem. And a practical joint transmission mode with relay selection and power allocation (JMRP) algorithm is developed to solve it. Simulation results show that the proposed algorithm outperforms the existing works in terms of system transmission energy and throughput.

The recently-emerged sparse representation in compressive sensing (CS) has gained extensive attention in signal processing and pattern recognition. In this paper, a new method of facial expression recognition based on Gabor wavelets and sparse representation classifier (SRC) is presented. Gabor wavelets representations are firstly extracted to evaluate the performance of the SRC method on facial expression recognition tasks. Three representative classification methods, including artificial neural network (ANN), K-nearest neighbor (KNN), support vector machines (SVM), are used to compare with the SRC method. Experimental results on the popular JAFFE facial expression databases, demonstrate the promising performance of the presented SRC method on facial expression recognition tasks, outperforming the other used methods.

Kernel-based feature extraction is a popular new research direction in machine learning community. Considering the nonlinear manifold structure of speech data, in this paper a recently proposed kernel manifold learning method, called kernel isometric mapping (KIsomap), is adopted as a mechanism for feature extraction on spoken emotion recognition tasks. KIsomap is used to extract the low-dimensional embedded feature data from original high-dimensional emotional acoustic features for spoken emotion recognition. Experimental results on the popular emotional German Berlin speech corpus show that KIsomap achieves better performance than other used two typical manifold learning methods, i.e., locally linear embedding (LLE) and isometric mapping (Isomap).

Spoken emotion recognition is an interesting and challenging subject. In this paper, a new feature extraction method based on local Fisher discriminant analysis (LFDA) is proposed for spoken emotion recognition. LFDA is used to extract the low-dimensional discriminant embedded feature data from high-dimensional emotional speech features on spoken emotion recognition tasks. The performance of LFDA is compared with principal component analysis (PCA) and linear discriminant analysis (LDA). Experimental results on the emotional Chinese speech database demonstrate the promising performance of the proposed method.

This paper introduces the principle of image scrambling and information hiding, and proposes a double random scrambling based on image blocks. The image information hiding and extraction system is implemented using Matlab, and its resistance to cropping and scribbling is analyzed. The proposed algorithm has the advantages of high efficiency and simple implementation.

Recently manifold learning algorithms researches have been motivated by the idea of modelling high dimensional data using approximate low dimensional submanifold of the original space. In this paper, two manifold learning algorithms, locally linear embedding (LLE) and isometric feature mapping (Isomap), are proposed to apply to speech phoneme feature data extracted from TIMIT corpus in an effort to perform nonlinear dimensionality reduction for yielding low dimensional features capable of discriminating between phonemes. Compared with these manifold learning algorithms, the traditional principal component analysis (PCA) is also used to perform linear dimensionality reduction within speech phoneme feature data. The resulting features are evaluated in support vector machines (SVM)-based phoneme recognition experiments. Experiment results indicate that manifold learning algorithms are effective for identifying phoneme using low dimensional phoneme feature data obtained from the original high dimensional phoneme feature space.

In this paper, linear and nonlinear dimensionality reduction algorithms are proposed to speech phoneme data from TIMIT corpus in an effort to perform dimensionality reduction for yielding low dimensional features capable of discriminating between phonemes. The linear dimensionality reduction method, including principal component analysis (PCA) and linear discriminant analysis (LDA), and the nonlinear dimensionality reduction method, including locally linear embedding (LLE) and isometric feature mapping (Isomap), are investigated. The resulting features by dimensionality reduction are evaluated in support vector machines (SVM)-based phoneme recognition experiments. Experiment results indicate that traditional linear LDA and PCA techniques for dimensionality reduction are capable of outperforming nonlinear LLE and Isomap techniques for phoneme recognition.

Nowadays, recognizing human emotion in speech signal has attracted much attention and plays an important role in affect computing, artificial intelligence and signal processing areas. In this paper, seven feature selection filtering methods, including CFS, Chisquare, consistency, gain ratio, Infogain, relief and symmetrical uncertainty, are proposed to perform feature selection from 47 original features extracted from natural emotional speech corpus in an effort to yield less features capable of discriminating between emotion categories. The results of feature selection were evaluated through a simple k-nearest-neighbors classifier. Experiment results indicate that the most important two features among all extracted features for emotion recognition are ratio of voiced to unvoiced frames and intensity max, achieving acceptable emotion recognition rate of 67%. Additionally, the features selected by Relief can obtain the highest accuracy of 72% with 9 features, outperforming all other feature selection methods and exceeding the mean accuracy of 71.5% with all features.

A new method of speech emotion recognition in speech signal via Fuzzy Least Squares Support Vector Machines (FLSSVM) is proposed for speech emotion recognition. Based on extracting prosody and voice quality features from emotional speech, FLSSVM is used to construct the optimum separating hyperplane to realize recognizing the four main speech emotion in Chinese including anger, happiness, sadness and surprise. Compared with other present methods of speech emotion recognition, computer simulation results show that FLSSVM can achieve higher average correct rate and better anti-noise recognition effect in different level of signal-to-noise ratios. This demonstrates the efficiency of the proposed FLSSVM method.

In this paper a new texture-based change detection approach is proposed to identify the flooded regions in SAR images. The main novelty of our approach is that the most distinctive texture information is automatically learned from the training set. Forty texture features, which are extracted from a pair of bi-temporal SAR images, are used to construct the weak classifier pool. After AdaBoost training, a strong classifier is optimally combined by a small subset of the candidate weak classifiers. The experimental results demonstrate the effectiveness of the proposed approach.