arXiv subjects

Dong Wang

Publications and source records attributed to Dong Wang.

At least 775 records · Page 43Linked to original sources

arXiv2016

Max-margin Metric Learning for Speaker Recognition

Probabilistic linear discriminant analysis (PLDA) is a popular normalization approach for the i-vector model, and has delivered state-of-the-art performance in speaker recognition. A potential problem of the PLDA model, however, is that it essentially assumes Gaussian distributions over speaker vectors, which is not always true in practice. Additionally, the objective function is not directly related to the goal of the task, e.g., discriminating true speakers and imposters. In this paper, we propose a max-margin metric learning approach to solve the problems. It learns a linear transform with a criterion that the margin between target and imposter trials are maximized. Experiments conducted on the SRE08 core test show that compared to PLDA, the new approach can obtain comparable or even better performance, though the scoring is simply a cosine computation.

cs.SD↗

arXiv2016

Nonintersecting Brownian motions on the unit circle

We consider an ensemble of $n$ nonintersecting Brownian particles on the unit circle with diffusion parameter $n^{-1/2}$, which are conditioned to begin at the same point and to return to that point after time $T$, but otherwise not to intersect. There is a critical value of $T$ which separates the subcritical case, in which it is vanishingly unlikely that the particles wrap around the circle, and the supercritical case, in which particles may wrap around the circle. In this paper, we show that in the subcritical and critical cases the probability that the total winding number is zero is almost surely 1 as $n\to\infty$, and in the supercritical case that the distribution of the total winding number converges to the discrete normal distribution. We also give a streamlined approach to identifying the Pearcey and tacnode processes in scaling limits. The formula of the tacnode correlation kernel is new and involves a solution to a Lax system for the Painlevé II equation of size 2 $\times$ 2. The proofs are based on the determinantal structure of the ensemble, asymptotic results for the related system of discrete Gaussian orthogonal polynomials, and a formulation of the correlation kernel in terms of a double contour integral.

math.PR↗

arXiv2016

A Universal Update-pacing Framework For Visual Tracking

This paper proposes a novel framework to alleviate the model drift problem in visual tracking, which is based on paced updates and trajectory selection. Given a base tracker, an ensemble of trackers is generated, in which each tracker's update behavior will be paced and then traces the target object forward and backward to generate a pair of trajectories in an interval. Then, we implicitly perform self-examination based on trajectory pair of each tracker and select the most robust tracker. The proposed framework can effectively leverage temporal context of sequential frames and avoid to learn corrupted information. Extensive experiments on the standard benchmark suggest that the proposed framework achieves superior performance against state-of-the-art trackers.

cs.CV↗

arXiv2015

Relation Classification via Recurrent Neural Network

Deep learning has gained much success in sentence-level relation classification. For example, convolutional neural networks (CNN) have delivered competitive performance without much effort on feature engineering as the conventional pattern-based methods. Thus a lot of works have been produced based on CNN structures. However, a key issue that has not been well addressed by the CNN-based method is the lack of capability to learn temporal features, especially long-distance dependency between nominal pairs. In this paper, we propose a simple framework based on recurrent neural networks (RNN) and compare it with CNN-based model. To show the limitation of popular used SemEval-2010 Task 8 dataset, we introduce another dataset refined from MIMLRE(Angeli et al., 2014). Experiments on two different datasets strongly indicates that the RNN-based model can deliver better performance on relation classification, and it is particularly capable of learning long-distance relation patterns. This makes it suitable for real-world applications where complicated expressions are often involved.

cs.CL↗

arXiv2015

Wakefield issue and its impact on X-ray photon pulse in the SXFEL test facility

Besides the designed beam acceleration, the energy of electrons changed by the longitudinal wakefields in a real free-electron laser (FEL) facility, which may degrade FEL performances from the theoretical expectation. In this paper, with the help of simulation codes, the wakefields induced beam energy loss in the sophisticated undulator section is calculated for Shanghai soft X-ray FEL, which is a two-stage seeded FEL test facility. While the 1st stage 44 nm FEL output is almost not affected by the wakefields, it is found that a beam energy loss about 0.8 MeV degrades the peak brightness of the 2nd stage 8.8 nm FEL by a factor of 1.6, which however can be compensated by a magnetic field fine tuning of each undulator segment.

physics.acc-ph↗

arXiv2015

THCHS-30 : A Free Chinese Speech Corpus

Speech data is crucially important for speech recognition research. There are quite some speech databases that can be purchased at prices that are reasonable for most research institutes. However, for young people who just start research activities or those who just gain initial interest in this direction, the cost for data is still an annoying barrier. We support the `free data' movement in speech recognition: research institutes (particularly supported by public funds) publish their data freely so that new researchers can obtain sufficient data to kick of their career. In this paper, we follow this trend and release a free Chinese speech database THCHS-30 that can be used to build a full- edged Chinese speech recognition system. We report the baseline system established with this database, including the performance under highly noisy conditions.

cs.CL↗

arXiv2015

Joint Source Selection and Data Extrapolation in Social Sensing for Disaster Response

This paper complements the large body of social sensing literature by developing means for augmenting sensing data with inference results that "fill-in" missing pieces. It specifically explores the synergy between (i) inference techniques used for filling-in missing pieces and (ii) source selection techniques used to determine which pieces to retrieve in order to improve inference results. We focus on prediction in disaster scenarios, where disruptive trend changes occur. We first discuss our previous conference study that compared a set of prediction heuristics and developed a hybrid prediction algorithm. We then enhance the prediction scheme by considering algorithms for sensor selection that improve inference quality. Our proposed source selection and extrapolation algorithms are tested using data collected during the New York City crisis in the aftermath of Hurricane Sandy in November 2012. The evaluation results show that consistently good predictions are achieved. The work is notable for addressing the bi-modal nature of damage propagation in complex systems subjected to stress, where periods of calm are interspersed with periods of severe change. It is novel in offering a new solution to the problem that jointly leverages source selection and extrapolation components thereby improving the results.

cs.SI↗

arXiv2015

Transfer Learning for Speech and Language Processing

Transfer learning is a vital technique that generalizes models trained for one setting or task to other settings or tasks. For example in speech recognition, an acoustic model trained for one language can be used to recognize speech in another language, with little or no re-training data. Transfer learning is closely related to multi-task learning (cross-lingual vs. multilingual), and is traditionally studied in the name of `model adaptation'. Recent advance in deep learning shows that transfer learning becomes much easier and more effective with high-level abstract features learned by deep models, and the `transfer' can be conducted not only between data distributions and data types, but also between model structures (e.g., shallow nets and deep nets) or even model types (e.g., Bayesian models and neural models). This review paper summarizes some recent prominent research towards this direction, particularly for speech and language processing. We also report some results from our group and highlight the potential of this very interesting research field.

cs.CL↗

arXiv2015

Deep Representation of Facial Geometric and Photometric Attributes for Automatic 3D Facial Expression Recognition

In this paper, we present a novel approach to automatic 3D Facial Expression Recognition (FER) based on deep representation of facial 3D geometric and 2D photometric attributes. A 3D face is firstly represented by its geometric and photometric attributes, including the geometry map, normal maps, normalized curvature map and texture map. These maps are then fed into a pre-trained deep convolutional neural network to generate the deep representation. Then the facial expression prediction is simplyachieved by training linear SVMs over the deep representation for different maps and fusing these SVM scores. The visualizations show that the deep representation provides a complete and highly discriminative coding scheme for 3D faces. Comprehensive experiments on the BU-3DFE database demonstrate that the proposed deep representation can outperform the widely used hand-crafted descriptors (i.e., LBP, SIFT, HOG, Gabor) and the state-of-art approaches under the same experimental protocols.

cs.CV↗

arXiv2015

Self-organized magnetic particles to tune the mechanical behaviour of a granular system

Above a certain density a granular material jams. This property can be controlled by either tuning a global property, such as the packing fraction or by applying shear strain, or at the micro-scale by tuning grain shape, inter-particle friction or externally controlled organization. Here, we introduce a novel way to change a local granular property by adding a weak anisotropic magnetic interaction between particles. We measure the evolution of the pressure, $P$, and coordination number, $Z$, for a packing of 2D photo-elastic disks, subject to uniaxial compression. Some of the particles have embedded cuboidal magnets. The strength of the magnetic interactions between particles are too weak to have a strong direct effect on $P$ or $Z$ when the system is jammed. However, the magnetic interactions play an important role in the evolution of latent force networks when systems containing a large enough fraction of the particles with magnets are driven through unjammed states. In this case, a statistically stable network of magnetic chains self-organizes and overlaps with force chains, strengthening the granular medium. We believe this property can be used to reversibly control mechanical properties of granular materials.

cond-mat.soft↗

arXiv2015

Stochastic Top-k ListNet

ListNet is a well-known listwise learning to rank model and has gained much attention in recent years. A particular problem of ListNet, however, is the high computation complexity in model training, mainly due to the large number of object permutations involved in computing the gradients. This paper proposes a stochastic ListNet approach which computes the gradient within a bounded permutation subset. It significantly reduces the computation complexity of model training and allows extension to Top-k models, which is impossible with the conventional implementation based on full-set permutations. Meanwhile, the new approach utilizes partial ranking information of human labels, which helps improve model quality. Our experiments demonstrated that the stochastic ListNet method indeed leads to better ranking performance and speeds up the model training remarkably.

cs.IR↗

arXiv2015

Correlation kernels for sums and products of random matrices

Let $X$ be a random matrix whose squared singular value density is a polynomial ensemble. We derive double contour integral formulas for the correlation kernels of the squared singular values of $GX$ and $TX$, where $G$ is a complex Ginibre matrix and $T$ is a truncated unitary matrix. We also consider the product of $X$ and several complex Ginibre/truncated unitary matrices. As an application, we derive the precise condition for the squared singular values of the product of several truncated unitary matrices to follow a polynomial ensemble. We also consider the sum $H + M$ where $H$ is a GUE matrix and $M$ is a random matrix whose eigenvalue density is a polynomial ensemble. We show that the eigenvalues of $H + M$ follow a polynomial ensemble whose correlation kernel can be expressed as a double contour integral. As an application, we point out a connection to the two-matrix model.

math.PR↗

arXiv2015

Fluctuations of TASEP and LPP with general initial data

We prove Airy process variational formulas for the one-point probability distribution of (discrete time parallel update) TASEP with general initial data, as well as last passage percolation from a general lattice path to a point. We also consider variants of last passage percolation with inhomogeneous parameter geometric weights and provide variational formulas of a similar nature. This proves one aspect of the conjectural description of the renormalization fixed point of the Kardar-Parisi-Zhang universality class.

math.PR↗

arXiv2015

Learning from LDA using Deep Neural Networks

Latent Dirichlet Allocation (LDA) is a three-level hierarchical Bayesian model for topic inference. In spite of its great success, inferring the latent topic distribution with LDA is time-consuming. Motivated by the transfer learning approach proposed by~\newcite{hinton2015distilling}, we present a novel method that uses LDA to supervise the training of a deep neural network (DNN), so that the DNN can approximate the costly LDA inference with less computation. Our experiments on a document classification task show that a simple DNN can learn the LDA behavior pretty well, while the inference is speeded up tens or hundreds of times.

cs.LG↗

arXiv2015

VMF-SNE: Embedding for Spherical Data

T-SNE is a well-known approach to embedding high-dimensional data and has been widely used in data visualization. The basic assumption of t-SNE is that the data are non-constrained in the Euclidean space and the local proximity can be modelled by Gaussian distributions. This assumption does not hold for a wide range of data types in practical applications, for instance spherical data for which the local proximity is better modelled by the von Mises-Fisher (vMF) distribution instead of the Gaussian. This paper presents a vMF-SNE embedding algorithm to embed spherical data. An iterative process is derived to produce an efficient embedding. The results on a simulation data set demonstrated that vMF-SNE produces better embeddings than t-SNE for spherical data.

cs.LG↗

arXiv2015

Nanoscale $β$-Nuclear Magnetic Resonance Depth Imaging of Topological Insulators

Considerable evidence suggests that variations in the properties of topological insulators (TIs) at the nanoscale and at interfaces can strongly affect the physics of topological materials. Therefore, a detailed understanding of surface states and interface coupling is crucial to the search for and applications of new topological phases of matter. Currently, no methods can provide depth profiling near surfaces or at interfaces of topologically inequivalent materials. Such a method could advance the study of interactions. Herein we present a non-invasive depth-profiling technique based on $β$-NMR spectroscopy of radioactive $^8$Li$^+$ ions that can provide "one-dimensional imaging" in films of fixed thickness and generates nanoscale views of the electronic wavefunctions and magnetic order at topological surfaces and interfaces. By mapping the $^8$Li nuclear resonance near the surface and 10 nm deep into the bulk of pure and Cr-doped bismuth antimony telluride films, we provide signatures related to the TI properties and their topological non-trivial characteristics that affect the electron-nuclear hyperfine field, the metallic shift and magnetic order. These nanoscale variations in $β$-NMR parameters reflect the unconventional properties of the topological materials under study, and understanding the role of heterogeneities is expected to lead to the discovery of novel phenomena involving quantum materials.

cond-mat.mtrl-sci↗

arXiv2015

Improved Deep Speaker Feature Learning for Text-Dependent Speaker Recognition

A deep learning approach has been proposed recently to derive speaker identifies (d-vector) by a deep neural network (DNN). This approach has been applied to text-dependent speaker recognition tasks and shows reasonable performance gains when combined with the conventional i-vector approach. Although promising, the existing d-vector implementation still can not compete with the i-vector baseline. This paper presents two improvements for the deep learning approach: a phonedependent DNN structure to normalize phone variation, and a new scoring approach based on dynamic time warping (DTW). Experiments on a text-dependent speaker recognition task demonstrated that the proposed methods can provide considerable performance improvement over the existing d-vector implementation.

cs.CL↗

arXiv2015

Recognize Foreign Low-Frequency Words with Similar Pairs

Low-frequency words place a major challenge for automatic speech recognition (ASR). The probabilities of these words, which are often important name entities, are generally under-estimated by the language model (LM) due to their limited occurrences in the training data. Recently, we proposed a word-pair approach to deal with the problem, which borrows information of frequent words to enhance the probabilities of low-frequency words. This paper presents an extension to the word-pair method by involving multiple `predicting words' to produce better estimation for low-frequency words. We also employ this approach to deal with out-of-language words in the task of multi-lingual speech recognition.

cs.CL↗