arXiv subjects

Yang Song

Publications and source records attributed to Yang Song.

At least 361 records · Page 20Linked to original sources

arXiv2017

Learning Unified Embedding for Apparel Recognition

In apparel recognition, specialized models (e.g. models trained for a particular vertical like dresses) can significantly outperform general models (i.e. models that cover a wide range of verticals). Therefore, deep neural network models are often trained separately for different verticals. However, using specialized models for different verticals is not scalable and expensive to deploy. This paper addresses the problem of learning one unified embedding model for multiple object verticals (e.g. all apparel classes) without sacrificing accuracy. The problem is tackled from two aspects: training data and training difficulty. On the training data aspect, we figure out that for a single model trained with triplet loss, there is an accuracy sweet spot in terms of how many verticals are trained together. To ease the training difficulty, a novel learning scheme is proposed by using the output from specialized models as learning targets so that L2 loss can be used instead of triplet loss. This new loss makes the training easier and make it possible for more efficient use of the feature space. The end result is a unified model which can achieve the same retrieval accuracy as a number of separate specialized models, while having the model complexity as one. The effectiveness of our approach is shown in experiments.

cs.CV↗

arXiv2017

Speed/accuracy trade-offs for modern convolutional object detectors

The goal of this paper is to serve as a guide for selecting a detection architecture that achieves the right speed/memory/accuracy balance for a given application and platform. To this end, we investigate various ways to trade accuracy for speed and memory usage in modern convolutional object detection systems. A number of successful systems have been proposed in recent years, but apples-to-apples comparisons are difficult due to different base feature extractors (e.g., VGG, Residual Networks), different default image resolutions, as well as different hardware and software platforms. We present a unified implementation of the Faster R-CNN [Ren et al., 2015], R-FCN [Dai et al., 2016] and SSD [Liu et al., 2015] systems, which we view as "meta-architectures" and trace out the speed/accuracy trade-off curve created by using alternative feature extractors and varying other critical parameters such as image size within each of these meta-architectures. On one extreme end of this spectrum where speed and memory are critical, we present a detector that achieves real time speeds and can be deployed on a mobile device. On the opposite end in which accuracy is critical, we present a detector that achieves state-of-the-art performance measured on the COCO detection task.

cs.CV↗

arXiv2017

Age Progression/Regression by Conditional Adversarial Autoencoder

"If I provide you a face image of mine (without telling you the actual age when I took the picture) and a large amount of face images that I crawled (containing labeled faces of different ages but not necessarily paired), can you show me what I would look like when I am 80 or what I was like when I was 5?" The answer is probably a "No." Most existing face aging works attempt to learn the transformation between age groups and thus would require the paired samples as well as the labeled query image. In this paper, we look at the problem from a generative modeling perspective such that no paired samples is required. In addition, given an unlabeled image, the generative model can directly produce the image with desired age attribute. We propose a conditional adversarial autoencoder (CAAE) that learns a face manifold, traversing on which smooth age progression and regression can be realized simultaneously. In CAAE, the face is first mapped to a latent vector through a convolutional encoder, and then the vector is projected to the face manifold conditional on age through a deconvolutional generator. The latent vector preserves personalized face features (i.e., personality) and the age condition controls progression vs. regression. Two adversarial networks are imposed on the encoder and generator, respectively, forcing to generate more photo-realistic faces. Experimental results demonstrate the appealing performance and flexibility of the proposed framework by comparing with the state-of-the-art and ground truth.

cs.CV↗

arXiv2017

Impurity-driven intervalley spin-flip scattering-induced 2D spin relaxation in silicon

Through the theoretical study of electron spin lifetime in the 2DEG of doped Si, we highlight a dominant spin relaxation mechanism induced by the impurity central-cell potential near an interface via intervalley electron scattering. At low temperatures and with modest doping, this Yafet spin flip mechanism can become much more important than the Dyakonov-Perel spin relaxation indcued by the structural Rashba spin-orbit coupling field. As the leading-order impurity-induced spin flip happens only between two non-opposite valleys in Si, two-dimensional electron gas (2DEG) systems in Si MOSFETs or SiGe heterostructures are a natural platform to test and utilize this spin relaxation mechanism due to the valley splitting near the interface and its tunability by electrical gating or applied stress. Our proposed new spin relaxation mechanism may explain a part of the spin relaxation contribution to Si-based 2DEG systems.

cond-mat.mes-hall↗

arXiv2017

Statistical exchange-coupling errors and the practicality of scalable silicon donor qubits

Recent experimental efforts have led to considerable interest in donor-based localized electron spins in Si as viable qubits for a scalable silicon quantum computer. With the use of isotopically purified $^{28}$Si and the realization of extremely long spin coherence time in single-donor electrons, the recent experimental focus is on two-coupled donors with the eventual goal of a scaled-up quantum circuit. Motivated by this development, we simulate the statistical distribution of the exchange coupling $J$ between a pair of donors under realistic donor placement straggles, and quantify the errors relative to the intended $J$ value. With $J$ values in a broad range of donor-pair separation ($5<|\mathbf{R}|<60$ nm), we work out various cases systematically, for a target donor separation $\mathbf{R}_0$ along the [001], [110] and [111] Si crystallographic directions, with $|\mathbf{R}_0|=10, 20$ or 30 nm and standard deviation $σ_R=1, 2, 5$ or 10 nm. Our extensive theoretical results demonstrate the great challenge for a prescribed $J$ gate even with just a donor pair, a first step for any scalable Si-donor-based quantum computer.

cond-mat.mes-hall↗

arXiv2016

Kernel Bayesian Inference with Posterior Regularization

We propose a vector-valued regression problem whose solution is equivalent to the reproducing kernel Hilbert space (RKHS) embedding of the Bayesian posterior distribution. This equivalence provides a new understanding of kernel Bayesian inference. Moreover, the optimization problem induces a new regularization for the posterior embedding estimator, which is faster and has comparable performance to the squared regularization in kernel Bayes' rule. This regularization coincides with a former thresholding approach used in kernel POMDPs whose consistency remains to be established. Our theoretical work solves this open problem and provides consistency analysis in regression settings. Based on our optimizational formulation, we propose a flexible Bayesian posterior regularization framework which for the first time enables us to put regularization at the distribution level. We apply this method to nonparametric state-space filtering tasks with extremely nonlinear dynamics and show performance gains over all other baselines.

stat.ML↗

arXiv2016

Derivative Delay Embedding: Online Modeling of Streaming Time Series

The staggering amount of streaming time series coming from the real world calls for more efficient and effective online modeling solution. For time series modeling, most existing works make some unrealistic assumptions such as the input data is of fixed length or well aligned, which requires extra effort on segmentation or normalization of the raw streaming data. Although some literature claim their approaches to be invariant to data length and misalignment, they are too time-consuming to model a streaming time series in an online manner. We propose a novel and more practical online modeling and classification scheme, DDE-MGM, which does not make any assumptions on the time series while maintaining high efficiency and state-of-the-art performance. The derivative delay embedding (DDE) is developed to incrementally transform time series to the embedding space, where the intrinsic characteristics of data is preserved as recursive patterns regardless of the stream length and misalignment. Then, a non-parametric Markov geographic model (MGM) is proposed to both model and classify the pattern in an online manner. Experimental results demonstrate the effectiveness and superior classification accuracy of the proposed DDE-MGM in an online setting as compared to the state-of-the-art.

cs.LG↗

arXiv2016

Suppressing spin relaxation in silicon

Uniaxial compressive strain along the [001] direction strongly suppresses the spin relaxation in silicon. When the strain level is large enough so that electrons are redistributed only in the two valleys along the strain axis, the dominant scattering mechanisms are quenched and electrons mainly experience intra-axis scattering processes (intravalley or intervalley scattering within valleys on the same crystal axis). We first derive the spin-flip matrix elements due to intra-axis electron scattering off impurities, and then provide a comprehensive model of the spin relaxation time due to all possible interactions of conduction-band electrons with impurities and phonons. We predict nearly three orders of magnitude improvement in the spin relaxation time of $\sim10^{19}\text{cm}^{-3}$ antimony-doped silicon (Si:Sb) at low temperatures.

cond-mat.mtrl-sci↗

arXiv2016

CHSalign: A Web Server That Builds upon Junction-Explorer and RNAJAG for Pairwise Alignment of RNA Secondary Structures with Coaxial Helical Stacking

RNA junctions are important structural elements of RNA molecules. They are formed when three or more helices come together in three-dimensional space. Recent studies have focused on the annotation and prediction of coaxial helical stacking (CHS) motifs within junctions. Here we exploit such predictions to develop an efficient alignment tool to handle RNA secondary structures with CHS motifs. Specifically, we build upon our Junction-Explorer software for predicting coaxial stacking and RNAJAG for modelling junction topologies as tree graphs to incorporate constrained tree matching and dynamic programming algorithms into a new method, called CHSalign, for aligning the secondary structures of RNA molecules containing CHS motifs. Thus, CHSalign is intended to be an efficient alignment tool for RNAs containing similar junctions. Experimental results based on thousands of alignments demonstrate that CHSalign can align two RNA secondary structures containing CHS motifs more accurately than other RNA secondary structure alignment tools. CHSalign yields a high score when aligning two RNA secondary structures with similar CHS motifs or helical arrangement patterns, and a low score otherwise. This new method has been implemented in a web server, and the program is also made freely available, at http://bioinformatics.njit.edu/CHSalign/.

cs.CE↗

arXiv2016

Fast control of semiconductor qubits beyond the rotating-wave approximation

We present a theoretical study of single-qubit operations by oscillatory fields on various semiconductor platforms. We explicitly show how to perform faster gate operations by going beyond the universally-used rotating wave approximation (RWA) regime, while using only two sinusoidal pulses. No complicated pulse shaping or optimal control sequences are required. We first show for specific published experiments how much error is currently incurred by implementing pulses designed using standard RWA. We then show that an even modest increase in gate speed would cause problems in using RWA for gate design in the singlet-triplet (ST) and resonant-exchange (RX) qubits. We discuss the extent to which analytically keeping higher orders in the perturbation theory would address the problem. More strikingly, we give a new prescription for gating with strong coupling far beyond the RWA regime. We perform numerical calculations for the phases and the durations of two consecutive pulses to realize the key Hadamard and $\fracπ{8}$ gates with coupling strengths up to several times the qubit splitting. Working in this manifestly non-RWA regime, the gate operation speeds up by two to three orders of magnitude.

cond-mat.mes-hall↗

arXiv2016

Training Deep Neural Networks via Direct Loss Minimization

Supervised training of deep neural nets typically relies on minimizing cross-entropy. However, in many domains, we are interested in performing well on metrics specific to the application. In this paper we propose a direct loss minimization approach to train deep neural networks, which provably minimizes the application-specific loss function. This is often non-trivial, since these functions are neither smooth nor decomposable and thus are not amenable to optimization with standard gradient-based methods. We demonstrate the effectiveness of our approach in the context of maximizing average precision for ranking problems. Towards this goal, we develop a novel dynamic programming algorithm that can efficiently compute the weight updates. Our approach proves superior to a variety of baselines in the context of action classification and object detection, especially in the presence of label noise.

cs.LG↗

arXiv2016

Improving the Robustness of Deep Neural Networks via Stability Training

In this paper we address the issue of output instability of deep neural networks: small perturbations in the visual input can significantly distort the feature embeddings and output of a neural network. Such instability affects many deep architectures with state-of-the-art performance on a wide range of computer vision tasks. We present a general stability training method to stabilize deep networks against small input distortions that result from various types of common image processing, such as compression, rescaling, and cropping. We validate our method by stabilizing the state-of-the-art Inception architecture against these types of distortions. In addition, we demonstrate that our stabilized model gives robust state-of-the-art performance on large-scale near-duplicate detection, similar-image ranking, and classification on noisy datasets.

cs.CV↗

arXiv2016

Canonical correlation analysis of high-dimensional data with very small sample support

This paper is concerned with the analysis of correlation between two high-dimensional data sets when there are only few correlated signal components but the number of samples is very small, possibly much smaller than the dimensions of the data. In such a scenario, a principal component analysis (PCA) rank-reduction preprocessing step is commonly performed before applying canonical correlation analysis (CCA). We present simple, yet very effective approaches to the joint model-order selection of the number of dimensions that should be retained through the PCA step and the number of correlated signals. These approaches are based on reduced-rank versions of the Bartlett-Lawley hypothesis test and the minimum description length information-theoretic criterion. Simulation results show that the techniques perform well for very small sample sizes even in colored noise.

cs.IT↗

arXiv2015

Bayesian Matrix Completion via Adaptive Relaxed Spectral Regularization

Bayesian matrix completion has been studied based on a low-rank matrix factorization formulation with promising results. However, little work has been done on Bayesian matrix completion based on the more direct spectral regularization formulation. We fill this gap by presenting a novel Bayesian matrix completion method based on spectral regularization. In order to circumvent the difficulties of dealing with the orthonormality constraints of singular vectors, we derive a new equivalent form with relaxed constraints, which then leads us to design an adaptive version of spectral regularization feasible for Bayesian inference. Our Bayesian method requires no parameter tuning and can infer the number of latent factors automatically. Experiments on synthetic and real datasets demonstrate encouraging results on rank recovery and collaborative filtering, with notably good results for very sparse matrices.

math.NA↗

arXiv2015

Low-Complexity Channel Reconstruction Methods Based on SVD-ZF Precoding in Massive 3D-MIMO Systems

In this paper, we study the low-complexity channel reconstruction methods for downlink precoding in massive multiple-Input multiple-Output (MIMO) systems. When the user is allocated less streams than the number of its antennas, the base station (BS) or user usually utilizes the singular value decomposition (SVD) to get the effective channels, whose dimension is equal to the number of streams. This process is called channel reconstruction and done in BS for time division duplex (TDD) mode. However, with the increasing of antennas in BS, the computation burden of SVD is getting incredible. Here, we propose a series of novel low-complexity channel reconstruction methods for downlink precoding in 3D spatial channel model. We consider different correlations between elevation and azimuth antennas, and divide the large dimensional matrix SVD into two kinds of small-size matrix SVD. The simulation results show that the proposed methods only produce less than 10% float computation than the traditional SVD zero-forcing (SVD-ZF) precoding method, while keeping nearly the same performance of 1Gbps.

cs.IT↗

arXiv2015

S-index: Towards Better Metrics for Quantifying Research Impact

The ongoing growth in the volume of scientific literature available today precludes researchers from efficiently discerning the relevant from irrelevant content. Researchers are constantly interested in impactful papers, authors and venues in their respective fields. Moreover, they are interested in the so-called recent "rising stars" of these contexts which may lead to attractive directions for future work, collaborations or impactful publication venues. In this work, we address the problem of quantifying research impact in each of these contexts, in order to better direct attention of researchers and streamline the processes of comparison, ranking and evaluation of contribution. Specifically, we begin by outlining intuitive underlying assumptions that impact quantification methods should obey and evaluate when current state-of-the-art methods fail to satisfy these properties. To this end, we introduce the s-index metric which quantifies research impact through influence propagation over a heterogeneous citation network. s-index is tailored from these intuitive assumptions and offers a number of desirable qualities including robustness, natural temporality and straightforward extensibility from the paper impact to broader author and venue impact contexts. We evaluate its effectiveness on the publicly available Microsoft Academic Search citation graph with over 119 million papers and 1 billion citation edges with 103 million and 21 thousand associated authors and venues respectively.

cs.DL↗

arXiv2015

Polarization analysis of excitons in monolayer and bilayer transition-metal dichalcogenides

The polarization analysis of optical transitions in monolayer and bilayer transition-metal dichalcogenides provides invaluable information on the spin and valley (pseudospin) degrees of freedom. To explain optical properties of a given monolayer transition-metal dichalcogenide, one should consider (i) the order of its spin-split conduction bands, (ii) whether intervalley scattering is prone to phonon bottleneck, (iii) and whether valley mixing by electron-hole exchange can take place. Using these principles, we present a consistent physical picture that elucidates a variety of features in the optical spectra of these materials. We explain the differences between optical transitions in monolayer MoSe$_2$ and monolayer WSe$_2$, finding that indirect excitons in the latter correspond to several low-energy optical transitions that so far were attributed to excitons bound to impurities. A possible mechanism that can explain the vanishing polarization in MoSe$_2$ is discussed. Finally, we consider the effect of an out-of-plane electric field, showing that it can reduce the initial polarization of bright excitons due to a Rashba-type coupling with dark excitons.

cond-mat.mtrl-sci↗

arXiv2015

The Implementation of Hadoop-based Crawler System and Graphlite-based PageRank-Calculation In Search Engine

Nowadays, the size of the Internet is experiencing rapid growth. As of December 2014, the number of global Internet websites has more than 1 billion and all kinds of information resources are integrated together on the Internet, however,the search engine is to be a necessary tool for all users to retrieve useful information from vast amounts of web data. Generally speaking, a complete search engine includes the crawler system, index building systems, sorting systems and retrieval system. At present there are many open source implementation of search engine, such as lucene, solr, katta, elasticsearch, solandra and so on. The crawler system and sorting system is indispensable for any kind of search engine and in order to guarantee its efficiency, the former needs to update crawled vast amounts of data and the latter requires real-time to build index on newly crawled web pages and calculae its corresponding PageRank value. It is unlikely to accomplish such huge computation tasks depending on a single hardware implementation of the crawler system and sorting system,from which aspect, the distributed cluster technology is brought to the front. In this paper, we use the hadoop Map - Reduce computing framework to implement a distributed crawler system, and use the GraphLite, a distributed synchronous graph-computing framework, to achieve the real-time computation in getting the PageRank value of the new crawled web page.

cs.DC↗