Biologically-inspired stochastic vector matching for noise-robust information processing

“End of Moore’s Law” has recently become a topic. Keeping the signal-to-noise ratio (SNR) at the same level in the future will surely increase the energy density of smaller-sized transistors. Lowering the operating voltage will prevent this, but the SNR would inevitably degrade. Meanwhile, biological systems such as cells and brains possess robustness against noise in their information processing in spite of the strong influence of stochastic thermal noise. Inspired by the information processing of organisms, we propose a stochastic computing model to acquire information from noisy signals. Our model is based on vector matching, in which the similarities between the input vector carrying external noisy signals and the reference vectors prepared in advance as memorized templates are evaluated in a stochastic manner. This model exhibited robustness against the noise strength and its performance was improved by addition of noise with an appropriate strength, which is similar to a phenomenon observed in stochastic resonance. Because the stochastic vector matching we propose here has robustness against noise, it is a candidate for noisy information processing that is driven by stochastically-operating devices with low energy consumption in future. Moreover, the stochastic vector matching may be applied to memory-based information processing like that of the brain.     

A temporal hand gesture recognition system based on hog and motion trajectory

A real-time, rapid and robust gesture recognition system is usually hindered by difficulty of hand localization and complexity of hand gesture modeling, especially under complex background. For eliminating these obstacles, in this paper, we propose a method using histograms of oriented gradients features (HOG) and motion trajectory information for temporal hand gesture recognition in natural environment. We firstly localize hand in video stream based on hand detection by HOG and support vector machine algorithm (SVM). After hand localization, the motion trajectory information of consecutive hand gesture is extracted and a database of standard gestures is built. Finally, the Mahalanobis distance between input gesture and database is computed for recognition. As the experimental results shown, our method exhibits a good performance in real-time test.     

Fast Encoding Method for Image Vector Quantization Based on Multiple Appropriate Features to Estimate Euclidean Distance

The encoding process of finding the best-matched codeword (winner) for a certain input vector in image vector quantization (VQ) is computationally very expensive due to a lot of k-dimensional Euclidean distance computations. In order to speed up the VQ encoding process, it is beneficial to firstly estimate how large the Euclidean distance is between the input vector and a candidate codeword by using appropriate low dimensional features of a vector instead of an immediate Euclidean distance computation. If the estimated Euclidean distance is large enough, it implies that the current candidate codeword could not be a winner so that it can be rejected safely and thus avoid actual Euclidean distance computation. Sum (1-D), L₂ norm (1-D) and partial sums (2-D) of a vector are used together as the appropriate features in this paper because they are the first three simplest features. Then, four estimations of Euclidean distance between the input vector and a codeword are connected to each other by the Cauchy–Schwarz inequality to realize codeword rejection. For typical standard images with very different details (Lena, F-16, Pepper and Baboon), the final remaining must-do actual Euclidean distance computations can be eliminated obviously and the total computational cost including all overhead can also be reduced obviously compared to the state-of-the-art EEENNS method meanwhile keeping a full search (FS) equivalent PSNR.     

Switching non-local vector median filter

This paper describes a novel image filtering method that removes random-valued impulse noise superimposed on a natural color image. In impulse noise removal, it is essential to employ a switching-type filtering method, as used in the well-known switching median filter, to preserve the detail of an original image with good quality. In color image filtering, it is generally preferable to deal with the red (R), green (G), and blue (B) components of each pixel of a color image as elements of a vectorized signal, as in the well-known vector median filter, rather than as component-wise signals to prevent a color shift after filtering. By taking these fundamentals into consideration, we propose a switching-type vector median filter with non-local processing that mainly consists of a noise detector and a noise removal filter. Concretely, we propose a noise detector that proactively detects noise-corrupted pixels by focusing attention on the isolation tendencies of pixels of interest not in an input image but in difference images between RGB components. Furthermore, as the noise removal filter, we propose an extended version of the non-local median filter, we proposed previously for grayscale image processing, named the non-local vector median filter, which is designed for color image processing. The proposed method realizes a superior balance between the preservation of detail and impulse noise removal by proactive noise detection and non-local switching vector median filtering, respectively. The effectiveness and validity of the proposed method are verified in a series of experiments using natural color images.     

Uncertainty propagation in SEA for structural–acoustic coupled systems with non-deterministic parameters

Considering limited available information on uncertainties in structural - acoustic coupled systems, two methods namely the vertex method and the Legendre orthogonal polynomial based method for predicting their dynamic behavior are developed based on the Statistical Energy Analysis (SEA) approach. For the vertex method, an efficient program for determining coordinates of all vertices of the rectangular spanned by entries of the involved interval input vector is coded, which is well suited for an interval input vector in arbitrary dimension. Instead of calculating the extremum of the response of interest, a method for determining its minimal and maximal point vectors dimension by dimension with respect to uncertain parameters is proposed based on the Legendre orthogonal polynomial approximation. Following the theoretical analysis of the accuracy and efficiency of the proposed methods, their validation is performed by one numerical example and two applications.     

Multi-Hop Question Generation Using Hierarchical Encoding-Decoding and Context Switch Mechanism

Neural auto-regressive sequence-to-sequence models have been dominant in text generation tasks, especially the question generation task. However, neural generation models suffer from the global and local semantic semantic drift problems. Hence, we propose the hierarchical encoding–decoding mechanism that aims at encoding rich structure information of the input passages and reducing the variance in the decoding phase. In the encoder, we hierarchically encode the input passages according to its structure at four granularity-levels: [word, chunk, sentence, document]-level. Second, we progressively select the context vector from the document-level representations to the word-level representations at each decoding time step. At each time-step in the decoding phase, we progressively select the context vector from the document-level representations to word-level. We also propose the context switch mechanism that enables the decoder to use the context vector from the last step when generating the current word at each time-step.It provides a means of improving the stability of the text generation process during the decoding phase when generating a set of consecutive words. Additionally, we inject syntactic parsing knowledge to enrich the word representations. Experimental results show that our proposed model substantially improves the performance and outperforms previous baselines according to both automatic and human evaluation. Besides, we implement a deep and comprehensive analysis of generated questions based on their types.     

Sentiment Classification Method Based on Blending of Emoticons and Short Texts

With the development of Internet technology, short texts have gradually become the main medium for people to obtain information and communicate. Short text reduces the threshold of information production and reading by virtue of its short length, which is in line with the trend of fragmented reading in the context of the current fast-paced life. In addition, short texts contain emojis to make the communication immersive. However, short-text content means it contains relatively little information, which is not conducive to the analysis of sentiment characteristics. Therefore, this paper proposes a sentiment classification method based on the blending of emoticons and short-text content. Emoticons and short-text content are transformed into vectors, and the corresponding word vector and emoticon vector are connected into a sentencing matrix in turn. The sentence matrix is input into a convolution neural network classification model for classification. The results indicate that, compared with existing methods, the proposed method improves the accuracy of analysis.     

Infrared Small Target Detection Using PPCA

Probabilistic PCA (PPCA) is an extension of PCA which reformulated PCA in a probabilistic framework. In this paper we propose a infrared small target detection algorithm using PPCA analogous to the face detection scheme using PCA, or known as “eigenface”. By computing the parameters of PPCA, we map the input vector from the image onto a subspace. After reconstructing the vector, the distance between the original vector and the reconstructed one will indicate the possibility of the input being a target. Experimental results show the effectiveness of this algorithm compared with other methods.     

Densely Connected Neural Networks for Nonlinear Regression

Densely connected convolutional networks (DenseNet) behave well in image processing. However, for regression tasks, convolutional DenseNet may lose essential information from independent input features. To tackle this issue, we propose a novel DenseNet regression model where convolution and pooling layers are replaced by fully connected layers and the original concatenation shortcuts are maintained to reuse the feature. To investigate the effects of depth and input dimensions of the proposed model, careful validations are performed by extensive numerical simulation. The results give an optimal depth (19) and recommend a limited input dimension (under 200). Furthermore, compared with the baseline models, including support vector regression, decision tree regression, and residual regression, our proposed model with the optimal depth performs best. Ultimately, DenseNet regression is applied to predict relative humidity, and the outcome shows a high correlation with observations, which indicates that our model could advance environmental data science.     

Uncertainty relations based on Wigner-Yanase skew information

In this paper, we use certain norm inequalities to obtain new uncertain relations based on the Wigner-Yanase skew information. First for an arbitrary finite number of observables we derive an uncertainty relation outperforming previous lower bounds. We then propose new weighted uncertainty relations for two noncompatible observables. Two separable criteria via skew information are also obtained.     

Darboux Transformation for the Vector Sine-Gordon Equation and Integrable Equations on a Sphere

We propose a method for construction of Darboux transformations, which is a new development of the dressing method for Lax operators invariant under a reduction group. We apply the method to the vector sine-Gordon equation and derive its Bäcklund transformations. We show that there is a new Lax operator canonically associated with our Darboux transformation resulting an evolutionary differential-difference system on a sphere. The latter is a generalised symmetry for the chain of Bäcklund transformations. Using the re-factorisation approach and the Bianchi permutability of the Darboux transformations, we derive new vector Yang–Baxter map and integrable discrete vector sine-Gordon equation on a sphere.     

Unsupervised feature dimension reduction for classification of MR spectra

We present an unsupervised feature dimension reduction method for the classification of magnetic resonance spectra. The technique preserves spectral information, important for disease profiling. We propose to use this technique as a preprocessing step for computationally demanding wrapper-based feature subset selection. We show that the classification accuracy on an independent test set can be sustained while achieving considerable feature reduction. Our method is applicable to other classification techniques, such as neural networks, support vector machines, etc.     

Flexible generation of femtosecond cylindrical vector beams(Invited Paper)

Femtosecond(fs)cylindrical vector beams(CVBs)have found use in many applications in recent years.However,the existing rigid generation methods seriously limit its development.Here,we propose a flexible method for generating fs-CVBs with arbitrary polarization order by employing half wave plates and vortex retarders.The polarization state,autocorrelation width,pulse width,and spectrum features of the input and generated CVB pulses are measured and compared.The results verify that the generated CVBs remain in the fs regime with no appreciable temporal distortion,and the energy conversion efficiency can reach as high as 96.5%,even for a third-order beam.As a flexible way to generate fs-CVBs,this method will have great significance for many applications.     

Stochastic resonance in the FitzHugh-Nagumo model from a dynamic mutual information point of view

Stochastic resonance(SR) in a FitzHugh-Nagumo neuron model is investigated based on a dynamic mutual information (DMI) between the input and the corresponding output signals. The DMI is expressed in terms of the (cross)power spectra of the input and output time series. Both stochastic-periodic and aperiodic SR are treated based on the DMI and our results are in good accord with the SR measured by the signal to noise ratio(SNR) for the case of the stochastic-periodic input and the power norm for the case of the aperiodic input.     

基于综合特征的光纤周界安防系统高效入侵事件识别

在光纤周界安防系统中,急需对入侵事件实现准确而高效的识别,对事件特征做简练而恰当的描述是其关键所在.本文提出一种基于综合特征的入侵事件识别方法,该方法引入全相位滤波器组将输入信号并行分解为多个频率通道,以提取这些通道的归一化功率值;进而与信号过零率相结合,构成包含时域信息、频域信息的综合特征向量;最后将该特征向量馈入径向基函数神经网络即可准确识别出攀爬、敲击、晃动、剪切四种常见的入侵动作.实验证明,本文方法相比于现有的经验模态分解方法,不仅提高了精度,而且显著加快了识别速度.     

The Dirac Operator on Untrapped Surfaces

We establish a sharp extrinsic lower bound for the first eigenvalue of the Dirac operator of an untrapped surface in initial data sets without apparent horizon in terms of the norm of its mean curvature vector. The equality case leads to rigidity results for the constraint equations with spherical boundary as well as uniqueness results for surfaces with constant mean curvature vector field in Minkowski space.     

Ultrafast holographic Stokesmeter for polarization imaging in real time

We propose an ultrafast holographic Stokesmeter using a volume holographic substrate with two sets of two orthogonal gratings to identify all four Stokes parameters of the input beam. We derive the Mueller matrix of the proposed architecture and determine the constraints necessary for reconstructing the complete Stokes vector. The speed of this device is determined primarily by the channel spectral bandwidth (typically 100 GHz), corresponding to a few picoseconds.     

Iterative sparse reconstruction of spectral domain OCT signal

We propose and study an iterative sparse reconstruction for Fourier domain optical coherence tomography （FD OCT） image by solving a constrained optimization problem that minimizes L-1 norm. Our method takes the spectral shape of the OCT light source into consideration in the iterative image reconstruction procedure that allows deconvolution of the axial point spread function from the blurred image during reconstruction rather than after reconstruction. By minimizing the L-1 norm, the axial resolution and the signal to noise ratio of image can both be enhanced. The effectiveness of our method is validated using numerical simulation and experiment.     

Estimation of parameters of signals received by an antenna array

We propose a method for estimating signal parameters and the weight vector in adaptive antenna arrays (AA) based on estimation of the parameters of a minimum polynomial of a sample correlation matrix (CM) of the input process. Being highly efficient, this method allows us to estimate signal parameters with accuracies that are close to the Cramer-Rao limits.Translated from Izvestiya Vysshikh Uchebnykh Zavedenii, Radiofizika, Vol. 39, No. 9, pp. 1144–1160, September, 1996.     

New sub-grid stochastic acceleration model in LES of high-Reynolds-number flows

The experimental observations of intermittent dynamics of Lagrangian acceleration in a “free” high-Reynolds-number turbulence are shown to be consistent with the Kolmogorov-Oboukhov theory. In line with Kolmogorov-Oboukhov’s predictions, a new sub-grid scale (SGS) model is proposed and is combined with the Smagorinsky model. The new SGS model is focused on simulation of the non-resolved total acceleration vector by two stochastic processes: one for its norm, another for its direction. The norm is simulated by stochastic equation, which was derived from the log-normal stochastic process for turbulent kinetic energy dissipation rate, with the Reynolds number, as the parameter. The direction of the acceleration vector is suggested to be governed by random walk process, with correlation on the Kolmogorov’s timescale. In the framework of this model, a surrogate unfiltered velocity field is emulated by computation of the instantaneous model-equation. The coarse-grid computation of a high-Reynolds-number stationary homogeneous turbulence reproduced qualitatively the main intermittency effects, which were observed in experiment of ENS in Lyon. Contrary to the standard LES with the Smagorinsky eddy-viscosity model, the proposed model provided: (i) non-Gaussianity in the acceleration distribution with stretched tails; (ii) rapid decorrelation of acceleration vector components; (iii) “long memory” in correlation of its norm. The turbulent energy spectra of stationary and decaying homogeneous turbulence are also better predicted by the proposed model.