Attention-Based Temporal-Frequency Aggregation for Speaker Verification.

Convolutional neural networks (CNNs) have significantly promoted the development of speaker verification (SV) systems because of their powerful deep feature learning capability. In CNN-based SV systems, utterance-level aggregation is an important component, and it compresses the frame-level features generated by the CNN frontend into an utterance-level representation. However, most of the existing aggregation methods aggregate the extracted features across time and cannot capture the speaker-dependent information contained in the frequency domain. To handle this problem, this paper proposes a novel attention-based frequency aggregation method, which focuses on the key frequency bands that provide more information for utterance-level representation. Meanwhile, two more effective temporal-frequency aggregation methods are proposed in combination with the existing temporal aggregation methods. The two proposed methods can capture the speaker-dependent information contained in both the time domain and frequency domain of frame-level features, thus improving the discriminability of speaker embedding. Besides, a powerful CNN-based SV system is developed and evaluated on the TIMIT and Voxceleb datasets. The experimental results indicate that the CNN-based SV system using the temporal-frequency aggregation method achieves a superior equal error rate of 5.96% on Voxceleb compared with the state-of-the-art baseline models.

Neural Network-Based Autonomous Search Model with Undulatory Locomotion Inspired by Caenorhabditis Elegans.

Caenorhabditis elegans (C. elegans) exhibits sophisticated chemotaxis behavior with a unique locomotion pattern using a simple nervous system only and is, therefore, well suited to inspire simple, cost-effective robotic navigation schemes. Chemotaxis in C. elegans involves two complementary strategies: klinokinesis, which allows reorientation by sharp turns when moving away from targets; and klinotaxis, which gradually adjusts the direction of motion toward the preferred side throughout the movement. In this study, we developed an autonomous search model with undulatory locomotion that combines these two C. elegans chemotaxis strategies with its body undulatory locomotion. To search for peaks in environmental variables such as chemical concentrations and radiation in directions close to the steepest gradients, only one sensor is needed. To develop our model, we first evolved a central pattern generator and designed a minimal network unit with proprioceptive feedback to encode and propagate rhythmic signals; hence, we realized realistic undulatory locomotion. We then constructed adaptive sensory neuron models following real electrophysiological characteristics and incorporated a state-dependent gating mechanism, enabling the model to execute the two orientation strategies simultaneously according to information from a single sensor. Simulation results verified the effectiveness, superiority, and realness of the model. Our simply structured model exploits multiple biological mechanisms to search for the shortest-path concentration peak over a wide range of gradients and can serve as a theoretical prototype for worm-like navigation robots.

Deployment Optimization Method of Multistatic Radar for Constructing Circular Barrier Coverage.

To construct circular barrier coverage (CBC) with multistatic radars, a deployment optimization method based on equipartition strategy is proposed in this paper. In the method, the whole circular area is divided into several sub-circles with equal width, and each sub-circle is blanketed by a sub-CBC that is built based on the multistatic radar deployment patterns. To determine the optimal deployment patterns for each sub-CBC, the optimization conditions are firstly studied. Then, to optimize the deployment of the whole circular area, a model based on minimum deployment cost is proposed, and the proposed model is divided into two sub-models to solve the optimization issue. In the inner model, it is assumed that the width of a sub-circle is given. Based on the optimization conditions of the deployment pattern, integer linear programming (ILP) and exhaustive method (EM) are jointly adopted to determine the types and numbers of deployment patterns. Moreover, a modified formula is introduced to calculate the maximum valid number of receivers in a pattern, thus narrowing the search scope of the EM. In the outer model, the width of a sub-circle is assumed to be a variable, and the EM is adopted to determine the minimum total deployment cost and the optimal deployment patterns on each sub-circle. Moreover, the improved formula is exploited to determine the range of width for a sub-circle barrier and reduce the search scope of the EM. Finally, simulations are conducted in different conditions to verify the effectiveness of the proposed method. The simulation results indicate that the proposed method can spend less deployment cost and deploy fewer transmitters than the state-of-the-artwork.

Minimum Eigenvector Collaborative Representation Discriminant Projection for Feature Extraction.

High-dimensional signals, such as image signals and audio signals, usually have a sparse or low-dimensional manifold structure, which can be projected into a low-dimensional subspace to improve the efficiency and effectiveness of data processing. In this paper, we propose a linear dimensionality reduction method-minimum eigenvector collaborative representation discriminant projection-to address high-dimensional feature extraction problems. On the one hand, unlike the existing collaborative representation method, we use the eigenvector corresponding to the smallest non-zero eigenvalue of the sample covariance matrix to reduce the error of collaborative representation. On the other hand, we maintain the collaborative representation relationship of samples in the projection subspace to enhance the discriminability of the extracted features. Also, the between-class scatter of the reconstructed samples is used to improve the robustness of the projection space. The experimental results on the COIL-20 image object database, ORL, and FERET face databases, as well as Isolet database demonstrate the effectiveness of the proposed method, especially in low dimensions and small training sample size.

A Robust Nonrigid Point Set Registration Method Based on Collaborative Correspondences.

The nonrigid point set registration is one of the bottlenecks and has the wide applications in computer vision, pattern recognition, image fusion, video processing, and so on. In a nonrigid point set registration problem, finding the point-to-point correspondences is challengeable because of the various image degradations. In this paper, a robust method is proposed to accurately determine the correspondences by fusing the two complementary structural features, including the spatial location of a point and the local structure around it. The former is used to define the absolute distance (AD), and the latter is exploited to define the relative distance (RD). The AD-correspondences and the RD-correspondences can be established based on AD and RD, respectively. The neighboring corresponding consistency is employed to assign the confidence for each RD-correspondence. The proposed heuristic method combines the AD-correspondences and the RD-correspondences to determine the corresponding relationship between two point sets, which can significantly improve the corresponding accuracy. Subsequently, the thin plate spline (TPS) is employed as the transformation function. At each step, the closed-form solutions of the affine and nonaffine parts of TPS can be independently and robustly solved. It facilitates to analyze and control the registration process. Experimental results demonstrate that our method can achieve better performance than several existing state-of-the-art methods.

A Fast Algorithm for 2D DOA Estimation Using an Omnidirectional Sensor Array.

The traditional 2D MUSIC algorithm fixes the azimuth or the elevation, and searches for the other without considering the directions of sources. A spectrum peak diffusion effect phenomenon is observed and may be utilized to detect the approximate directions of sources. Accordingly, a fast 2D MUSIC algorithm, which performs azimuth and elevation simultaneous searches (henceforth referred to as AESS) based on only three rounds of search is proposed. Firstly, AESS searches along a circle to detect the approximate source directions. Then, a subsequent search is launched along several straight lines based on these approximate directions. Finally, the 2D Direction of Arrival (DOA) of each source is derived by searching on several small concentric circles. Unlike the 2D MUSIC algorithm, AESS does not fix any azimuth and elevation parameters. Instead, the adjacent point of each search possesses different azimuth and elevation, i.e., azimuth and elevation are simultaneously searched to ensure that the search path is minimized, and hence the total spectral search over the angular field of view is avoided. Simulation results demonstrate the performance characters of the proposed AESS over some existing algorithms.

RefSelect: a reference sequence selection algorithm for planted (l, d) motif search.

BACKGROUND: The planted (l, d) motif search (PMS) is an important yet challenging problem in computational biology. Pattern-driven PMS algorithms usually use k out of t input sequences as reference sequences to generate candidate motifs, and they can find all the (l, d) motifs in the input sequences. However, most of them simply take the first k sequences in the input as reference sequences without elaborate selection processes, and thus they may exhibit sharp fluctuations in running time, especially for large alphabets. RESULTS: In this paper, we build the reference sequence selection problem and propose a method named RefSelect to quickly solve it by evaluating the number of candidate motifs for the reference sequences. RefSelect can bring a practical time improvement of the state-of-the-art pattern-driven PMS algorithms. Experimental results show that RefSelect (1) makes the tested algorithms solve the PMS problem steadily in an efficient way, (2) particularly, makes them achieve a speedup of up to about 100× on the protein data, and (3) is also suitable for large data sets which contain hundreds or more sequences. CONCLUSIONS: The proposed algorithm RefSelect can be used to solve the problem that many pattern-driven PMS algorithms present execution time instability. RefSelect requires a small amount of storage space and is capable of selecting reference sequences efficiently and effectively. Also, the parallel version of RefSelect is provided for handling large data sets.

[Research on the interlaced encoding pixels in Hadamard transform spectral imager based on DMD].

The key innovation in Hadamard transform spectral imager (HTSI) introduced recently is the use of digital micro-mirror device (DMD) to encode spectral information. However, it brings some new problems for us to solve synchronously. An interlaced encoding phenomenon caused by the application of DMD to our HTSI was investigated and analyzed. These interlaced encoding pixels were not encoded based on Hadamard transform; therefore they should be processed specially in spectrum recovery. To improve the quality of the recovered spectral images, a positioning method and a decoding method for the interlaced encoding pixels were proposed. In our experiment, we first directed a beam of laser into our HTSI to fill the field of view and labeled the positions of the interlaced encoding pixels. Then we recorded two groups of the encoded images of the target by changing the positions of all the encoding channels on the DMD. The interlaced encoding pixels could be distinguished easily by observing the number of non-zero constants and zero elements in a column vector which is made up of the gray values of a pixel of the encoded images in sequence. The interlaced encoding pixels of the first group of the encoded images turned into the normal Hadamard encoding pixels of the second group of the encoded images. The interlaced encoding pixels of the first group of the encoded images can be decoded through applying inverse Hadamard transform to the corresponding pixels of the second group of the encoded images. The experimental results prove the feasibility of the decoding method.

Neural model generating klinotaxis behavior accompanied by a random walk based on C. elegans connectome.

Klinotaxis is a strategy of chemotaxis behavior in Caenorhabditis elegans (C. elegans), and random walking is evident during its locomotion. As yet, the understanding of the neural mechanisms underlying these behaviors has remained limited. In this study, we present a connectome-based simulation model of C. elegans to concurrently realize realistic klinotaxis and random walk behaviors and explore their neural mechanisms. First, input to the model is derived from an ASE sensory neuron model in which the all-or-none depolarization characteristic of ASEL neuron is incorporated for the first time. Then, the neural network is evolved by an evolutionary algorithm; klinotaxis emerged spontaneously. We identify a plausible mechanism of klinotaxis in this model. Next, we propose the liquid synapse according to the stochastic nature of biological synapses and introduce it into the model. Adopting this, the random walk is generated autonomously by the neural network, providing a new hypothesis as to the neural mechanism underlying the random walk. Finally, simulated ablation results are fairly consistent with the biological conclusion, suggesting the similarity between our model and the biological network. Our study is a useful step forward in behavioral simulation and understanding the neural mechanisms of behaviors in C. elegans.

High-precision synchronization detection method for bistatic radar.

To improve the accuracy of radar ranging and positioning under complex backgrounds, a high-precision synchronization detection method for bistatic radar is proposed based on different-frequency phase processing. First, the transmit signal and receive signal are converted to radio frequency pulses by frequency conversion. Then, the transmit signal is roughly measured with a field-programmable gate array. The obtained rough signal is used as the reference signal under the control of the direct digital synthesizer frequency synthesizer. Second, the transmit signal and receive signal are each synchronized with the reference signal in a different-frequency phase detection. These detection results are used as the starting signal and stopping signal of the counter gate. Finally, all the signals are counted, and the time synchronization between the transmit signal and the receive signal is implemented by processing the counting value. The experimental results show that a time synchronization precision of 1.5 ps and a frequency stability of 6.2 × 10-13/s can be reached. This method has the advantages of a fast time response, good noise suppression, and high measurement precision and strong system reliability.

PairMotifChIP: A Fast Algorithm for Discovery of Patterns Conserved in Large ChIP-seq Data Sets.

Identifying conserved patterns in DNA sequences, namely, motif discovery, is an important and challenging computational task. With hundreds or more sequences contained, the high-throughput sequencing data set is helpful to improve the identification accuracy of motif discovery but requires an even higher computing performance. To efficiently identify motifs in large DNA data sets, a new algorithm called PairMotifChIP is proposed by extracting and combining pairs of l-mers in the input with relatively small Hamming distance. In particular, a method for rapidly extracting pairs of l-mers is designed, which can be used not only for PairMotifChIP, but also for other DNA data mining tasks with the same demand. Experimental results on the simulated data show that the proposed algorithm can find motifs successfully and runs faster than the state-of-the-art motif discovery algorithms. Furthermore, the validity of the proposed algorithm has been verified on real data.

High-Resolution Group Quantization Phase Processing Method in Radio Frequency Measurement Range.

Aiming at the more complex frequency translation, the longer response time and the limited measurement precision in the traditional phase processing, a high-resolution phase processing method by group quantization higher than 100 fs level is proposed in radio frequency measurement range. First, the phase quantization is used as a step value to quantize every phase difference in a group by using the fixed phase relationships between different frequencies signals. The group quantization is formed by the results of the quantized phase difference. In the light of frequency drift mainly caused by phase noise of measurement device, a regular phase shift of the group quantization is produced, which results in the phase coincidence of two comparing signals which obtain high-resolution measurement. Second, in order to achieve the best coincidences pulse, a subtle delay is initiatively used to reduce the width of the coincidences fuzzy area according to the transmission characteristics of the coincidences in the specific medium. Third, a series of feature coincidences pulses of fuzzy area can be captured by logic gate to achieve the best phase coincidences information for the improvement of the measurement precision. The method provides a novel way to precise time and frequency measurement.

Neural network learning algorithms for tracking minor subspace in high-dimensional data stream.

A novel random-gradient-based algorithm is developed for online tracking the minor component (MC) associated with the smallest eigenvalue of the autocorrelation matrix of the input vector sequence. The five available learning algorithms for tracking one MC are extended to those for tracking multiple MCs or the minor subspace (MS). In order to overcome the dynamical divergence properties of some available random-gradient-based algorithms, we propose a modification of the Oja-type algorithms, called OJAm, which can work satisfactorily. The averaging differential equation and the energy function associated with the OJAm are given. It is shown that the averaging differential equation will globally asymptotically converge to an invariance set. The corresponding energy or Lyapunov functions exhibit a unique global minimum attained if and only if its state matrices span the MS of the autocorrelation matrix of a vector data stream. The other stationary points are saddle (unstable) points. The globally convergence of OJAm is also studied. The OJAm provides an efficient online learning for tracking the MS. It can track an orthonormal basis of the MS while the other five available algorithms cannot track any orthonormal basis of the MS. The performances of the relative algorithms are shown via computer simulations.

A neural network learning for adaptively extracting cross-correlation features between two high-dimensional data streams.

This paper proposes a novel cross-correlation neural network (CNN) model for finding the principal singular subspace of a cross-correlation matrix between two high-dimensional data streams. We introduce a novel nonquadratic criterion (NQC) for searching the optimum weights of two linear neural networks (LNN). The NQC exhibits a single global minimum attained if and only if the weight matrices of the left and right neural networks span the left and right principal singular subspace of a cross-correlation matrix, respectively. The other stationary points of the NQC are (unstable) saddle points. We develop an adaptive algorithm based on the NQC for tracking the principal singular subspace of a cross-correlation matrix between two high-dimensional vector sequences. The NQC algorithm provides a fast online learning of the optimum weights for two LNN. The global asymptotic stability of the NQC algorithm is analyzed. The NQC algorithm has several key advantages such as faster convergence, which is illustrated through simulations.

Image registration algorithm using Mexican hat function-based operator and grouped feature matching strategy.

Feature detection and matching are crucial for robust and reliable image registration. Although many methods have been developed, they commonly focus on only one class of image features. The methods that combine two or more classes of features are still novel and significant. In this work, methods for feature detection and matching are proposed. A Mexican hat function-based operator is used for image feature detection, including the local area detection and the feature point detection. For the local area detection, we use the Mexican hat operator for image filtering, and then the zero-crossing points are extracted and merged into the area borders. For the feature point detection, the Mexican hat operator is performed in scale space to get the key points. After the feature detection, an image registration is achieved by using the two classes of image features. The feature points are grouped according to a standardized region that contains correspondence to the local area, precise registration is achieved eventually by the grouped points. An image transformation matrix is estimated by the feature points in a region and then the best one is chosen through competition of a set of the transformation matrices. This strategy has been named the Grouped Sample Consensus (GCS). The GCS has also ability for removing the outliers effectively. The experimental results show that the proposed algorithm has high registration accuracy and small computational volume.

Adaptive quasi-Newton algorithm for source extraction via CCA approach.

This paper addresses the problem of adaptive source extraction via the canonical correlation analysis (CCA) approach. Based on Liu's analysis of CCA approach, we propose a new criterion for source extraction, which is proved to be equivalent to the CCA criterion. Then, a fast and efficient online algorithm using quasi-Newton iteration is developed. The stability of the algorithm is also analyzed using Lyapunov's method, which shows that the proposed algorithm asymptotically converges to the global minimum of the criterion. Simulation results are presented to prove our theoretical analysis and demonstrate the merits of the proposed algorithm in terms of convergence speed and successful rate for source extraction.

Adaptive improved natural gradient algorithm for blind source separation.

We propose an adaptive improved natural gradient algorithm for blind separation of independent sources. First, inspired by the well-known backpropagation algorithm, we incorporate a momentum term into the natural gradient learning process to accelerate the convergence rate and improve the stability. Then an estimation function for the adaptation of the separation model is obtained to adaptively control a step-size parameter and a momentum factor. The proposed natural gradient algorithm with variable step-size parameter and variable momentum factor is therefore particularly well suited to blind source separation in a time-varying environment, such as an abruptly changing mixing matrix or signal power. The expected improvement in the convergence speed, stability, and tracking ability of the proposed algorithm is demonstrated by extensive simulation results in both time-invariant and time-varying environments. The ability of the proposed algorithm to separate extremely weak or badly scaled sources is also verified. In addition, simulation results show that the proposed algorithm is suitable for separating mixtures of many sources (e.g., the number of sources is 10) in the complete case.