arXiv subjects

Ke Wang

Publications and source records attributed to Ke Wang.

At least 631 records · Page 35Linked to original sources

arXiv2018

Empirical Evaluation of Speaker Adaptation on DNN based Acoustic Model

Speaker adaptation aims to estimate a speaker specific acoustic model from a speaker independent one to minimize the mismatch between the training and testing conditions arisen from speaker variabilities. A variety of neural network adaptation methods have been proposed since deep learning models have become the main stream. But there still lacks an experimental comparison between different methods, especially when DNN-based acoustic models have been advanced greatly. In this paper, we aim to close this gap by providing an empirical evaluation of three typical speaker adaptation methods: LIN, LHUC and KLD. Adaptation experiments, with different size of adaptation data, are conducted on a strong TDNN-LSTM acoustic model. More challengingly, here, the source and target we are concerned with are standard Mandarin speaker model and accented Mandarin speaker model. We compare the performances of different methods and their combinations. Speaker adaptation performance is also examined by speaker's accent degree.

cs.SD↗

arXiv2018

AI Benchmark: Running Deep Neural Networks on Android Smartphones

Over the last years, the computational power of mobile devices such as smartphones and tablets has grown dramatically, reaching the level of desktop computers available not long ago. While standard smartphone apps are no longer a problem for them, there is still a group of tasks that can easily challenge even high-end devices, namely running artificial intelligence algorithms. In this paper, we present a study of the current state of deep learning in the Android ecosystem and describe available frameworks, programming models and the limitations of running AI on smartphones. We give an overview of the hardware acceleration resources available on four main mobile chipset platforms: Qualcomm, HiSilicon, MediaTek and Samsung. Additionally, we present the real-world performance results of different mobile SoCs collected with AI Benchmark that are covering all main existing hardware configurations.

cs.AI↗

arXiv2018

Impact of geometry and non-idealities on electron 'optics' based graphene p-n junction devices

We articulate the challenges and opportunities of unconventional devices using the photon like flow of electrons in graphene, such as Graphene Klein Tunnel (GKT) transistors. The underlying physics is the employment of momentum rather than energy filtering to engineer a gate tunable transport gap in a 2D Dirac cone bandstructure. In the ballistic limit, we get a clean tunable gap that implies subthermal switching voltages below the Boltzmann limit, while maintaining a high saturating current in the output characteristic. In realistic structures, detailed numerical simulations and experiments show that momentum scattering, especially from the edges, bleeds leakage paths into the transport gap and turns it into a pseudogap. We quantify the importance of reducing edge roughness and overall geometry on the low-bias transfer characteristics of GKT transistors and benchmark against experimental data. We find that geometry plays a critical role in determining the performance of electron optics based devices that utilize angular resolution of electrons.

cond-mat.mes-hall↗

arXiv2018

Graphene Transistor Based on Tunable Dirac-Fermion-Optics

The linear energy-momentum dispersion, coupled with pseudo-spinors, makes graphene an ideal solid-state material platform to realize an electronic device based on Dirac-Fermionic relativistic quantum mechanics. Employing local gate control, several examples of electronic devices based on Dirac fermion dynamics have been demonstrated, including Klein tunneling, negative refraction and specular Andreev reflection. In this work, we present a quantum switch based on analogous Dirac-fermion-optics (DFO), in which the angle dependence of Klein tunneling is explicitly utilized to build tunable collimators and reflectors for the quantum wave function of Dirac fermions. We employ a novel dual-source design with a single flat reflector, which minimizes diffusive edge scattering and suppresses the background incoherent transmission. Our gate-tunable collimator-reflector device design enables measurement of the net DFO contribution in the switching device operation. We measure a full set of transmission coefficients of DFO wavefunction between multiple leads of the device, separating the classical contribution from that of any disorder in the channel. Since the DFO quantum switch demonstrated in this work requires no explicit energy gap, the switching operation is expected to be robust against thermal fluctuations and inhomogeneity length scales comparable to the Fermi wavelength. We find our quantum switch works at an elevated temperature up to 230 K and large bias current density up to 102 A/m, over a wide range of carrier densities. The tunable collimator-reflector coupled with the conjugated source electrodes developed in this work provides an additional component to build more efficient DFO electronic devices.

cond-mat.mes-hall↗

arXiv2018

Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding

Top-$N$ sequential recommendation models each user as a sequence of items interacted in the past and aims to predict top-$N$ ranked items that a user will likely interact in a `near future'. The order of interaction implies that sequential patterns play an important role where more recent items in a sequence have a larger impact on the next item. In this paper, we propose a Convolutional Sequence Embedding Recommendation Model (\emph{Caser}) as a solution to address this requirement. The idea is to embed a sequence of recent items into an `image' in the time and latent spaces and learn sequential patterns as local features of the image using convolutional filters. This approach provides a unified and flexible network structure for capturing both general preferences and sequential patterns. The experiments on public datasets demonstrated that Caser consistently outperforms state-of-the-art sequential recommendation methods on a variety of common evaluation metrics.

cs.IR↗

arXiv2018

Ranking Distillation: Learning Compact Ranking Models With High Performance for Recommender System

We propose a novel way to train ranking models, such as recommender systems, that are both effective and efficient. Knowledge distillation (KD) was shown to be successful in image recognition to achieve both effectiveness and efficiency. We propose a KD technique for learning to rank problems, called \emph{ranking distillation (RD)}. Specifically, we train a smaller student model to learn to rank documents/items from both the training data and the supervision of a larger teacher model. The student model achieves a similar ranking performance to that of the large teacher model, but its smaller model size makes the online inference more efficient. RD is flexible because it is orthogonal to the choices of ranking models for the teacher and student. We address the challenges of RD for ranking problems. The experiments on public data sets and state-of-the-art recommendation models showed that RD achieves its design purposes: the student model learnt with RD has a model size less than half of the teacher model while achieving a ranking performance similar to the teacher model and much better than the student model learnt without RD.

cs.LG↗

arXiv2018

Considerations for utilizing sodium chloride in epitaxial molybdenum disulfide

The utilization of alkali salts, such as NaCl and KI, have enabled the successful growth of large single domain and fully coalesced polycrystalline two-dimensional (2D) transition metal dichalcogenide layers. However, the impact of alkali salts on photonic and electronic properties are not fully established. In this work, we report alkali-free epitaxy of MoS2 on sapphire and benchmark the properties against alkali-assisted growth of MoS2. This study demonstrates that although NaCl can dramatically increase the domain size of monolayer MoS2 by 20 times, it can also induce strong optical and electronic heterogeneities in as-grown large-scale films. This work elucidates that utilization of NaCl can lead to variation in growth rates, loss of epitaxy, and a high density of nanoscale MoS2 particles (4/μm2). Such phenomena suggest that alkali atoms play an important role in Mo and S adatom mobility and strongly influence the 2D/sapphire interface during growth. Compared to alkali-free synthesis under the same growth conditions, MoS2 growth assisted by NaCl results in >1% tensile strain in as-grown domains, which reduces photoluminescence by ~20x and degrades transistor performance.

cond-mat.mtrl-sci↗

arXiv2018

Reconstruction and Registration of Large-Scale Medical Scene Using Point Clouds Data from Different Modalities

Sensing the medical scenario can ensure the safety during the surgical operations. So, in this regard, a monitor platform which can obtain the accurate location information of the surgery room is desperately needed. Compared to 2D camera image, 3D data contains more information of distance and direction. Therefore, 3D sensors are more suitable to be used in surgical scene monitoring. However, each 3D sensor has its own limitations. For example, Lidar (Light Detection and Ranging) can detect large-scale environment with high precision, but the point clouds or depth maps are very sparse. As for commodity RGBD sensors, such as Kinect, can accurately capture denser data, but limited to a small range from 0.5 to 4.5m. So, a proper method which can address these problems for fusing different modalities data is important. In this paper, we proposed a method which can fuse different modalities 3D data to get a large-scale and dense point cloud. The key contributions of our work are as follows. First, we proposed a 3D data collecting system to reconstruct the medical scenes. By fusing the Lidar and Kinect data, a large-scale medical scene with more details can be reconstructed. Second, we proposed a location-based fast point clouds registration algorithm to deal with different modality datasets.

cs.CV↗

arXiv2018

Distance Priors from Planck Final Release

We present the distance priors from the finally released $Planck~ \textrm{TT,TE,EE}+\textrm{lowE}$ data in 2018. The uncertainties are around $40\%$ smaller than those from $Planck$ 2015 TT$+$lowP. In order to check the validity of these new distance priors, we adopt the distance priors to constrain the cosmological parameters in different dark energy models, including the $Λ$CDM model, the $w$CDM model and the CPL model, and conclude that the distance priors provide consistent constraints on the relevant cosmological parameters compared to those from the full $Planck$ 2018 data release.

astro-ph.CO↗

arXiv2018

Zooming in to Massive Star Birth

We present high resolution (0.2", 1000 AU) 1.3 mm ALMA observations of massive infrared dark cloud clump, G028.37+00.07-C1, thought to harbor the early stages of massive star formation. Using $\rm N_2D^+$(3-2) we resolve the previously identified C1-S core, separating the bulk of its emission from two nearby protostellar sources. C1-S is thus identified as a massive ($\sim50\:M_\odot$), compact ($\sim0.1\:$pc diameter) starless core, e.g., with no signs of outflow activity. Being highly deuterated, this is a promising candidate for a pre-stellar core on the verge of collapse. An analysis of its dynamical state indicates a sub-virial velocity dispersion compared to a trans-Alfvénic turbulent core model. However, virial equilibrium could be achieved with sub-Alfvénic conditions involving $\sim2\:$mG magnetic field strengths.

astro-ph.GA↗

arXiv2018

Numerical Relativity Investigation of the Effects of Gravitational Waves on the Inhomogeneity of the Universe

We numerically integrate the Einstein's equations for a spatially flat Friedmann-Lemaire-Robertson-Walker (FLRW) background spacetime with a spatial curvature perturbation and evolving primordial tensor perturbations using the Einstein Toolkit. We find that although the primordial tensor perturbation doesn't play an important role in the evolution of the overdensity produced by the scalar perturbation, there is an obvious imprint left by the primordial tensor perturbation on the distribution of the fractional density perturbation in the nonlinear region. This imprint may be a possible probe of a gravitational waves background in the future.

astro-ph.CO↗

arXiv2018

Dynamic Neural Program Embedding for Program Repair

Neural program embeddings have shown much promise recently for a variety of program analysis tasks, including program synthesis, program repair, fault localization, etc. However, most existing program embeddings are based on syntactic features of programs, such as raw token sequences or abstract syntax trees. Unlike images and text, a program has an unambiguous semantic meaning that can be difficult to capture by only considering its syntax (i.e. syntactically similar pro- grams can exhibit vastly different run-time behavior), which makes syntax-based program embeddings fundamentally limited. This paper proposes a novel semantic program embedding that is learned from program execution traces. Our key insight is that program states expressed as sequential tuples of live variable values not only captures program semantics more precisely, but also offer a more natural fit for Recurrent Neural Networks to model. We evaluate different syntactic and semantic program embeddings on predicting the types of errors that students make in their submissions to an introductory programming class and two exercises on the CodeHunt education platform. Evaluation results show that our new semantic program embedding significantly outperforms the syntactic program embeddings based on token sequences and abstract syntax trees. In addition, we augment a search-based program repair system with the predictions obtained from our se- mantic embedding, and show that search efficiency is also significantly improved.

cs.AI↗

arXiv2018

Enhanced Robustness of Zero-line Modes in Graphene via a Magnetic Field

Motivated by recent experimental results for zero-line modes (ZLMs) in a bilayer graphene system [Nature Nanotechnol. 11, 1060 (2016)], we systematically studied the influence of a magnetic field on ZLMs and demonstrated the physical origin of the enhanced robustness by employing nonequilibrium Green's functions and the Landauer-Buttiker formula. We found that a perpendicular magnetic field can separate the wavefunctions of the counter-propagating kink states into opposite directions. Specifically, the separation vanishes at the charge neutrality point. The separation increases as the Fermi level deviates from the charge neutrality point and can reach a magnitude comparable to the wavefunction spread at a moderate field strength. Such spatial separation of oppositely propagating ZLMs effectively suppresses backscattering. Moreover, the presence of a magnetic field enlarges the bulk gap and suppresses the bound states, thereby further reducing the scattering. These mechanisms effectively increase the mean free paths of the ZLMs to approximately 1 micron in the presence of a disorder.

cond-mat.mes-hall↗

arXiv2018

A holistic perspective on the dynamics of G035.39-00.33: the interplay between gas and magnetic fields

Magnetic field is one of the key agents that play a crucial role in shaping molecular clouds and regulating star formation, yet the complete information on the magnetic field is not well constrained due to the limitations in observations. We study the magnetic field in the massive infrared dark cloud G035.39-00.33 from dust continuum polarization observations at 850 $\micron$ with SCUBA-2/POL-2 at JCMT. The magnetic field tends to be perpendicular to the densest part of the main filament (F$_{M}$), whereas it has a less defined relative orientation in the rest of the structure, where it tends to be parallel to some diffuse regions. A mean plane-of-the-sky magnetic field strength of $\sim$50 $μ$G for F$_{M}$ is obtained using Davis-Chandrasekhar-Fermi method. Based on $^{13}$CO (1-0) line observations, we suggest a formation scenario of F$_{M}$ due to large-scale ($\sim$10 pc) cloud-cloud collision. Using additional NH$_3$ line data, we estimate that F$_{M}$ will be gravitationally unstable if it is only supported by thermal pressure and turbulence. The northern part of F$_{M}$, however, can be stabilized by a modest additional support from the local magnetic field. The middle and southern parts of F$_{M}$ are likely unstable even if the magnetic field support is taken into account. We claim that the clumps in F$_{M}$ may be supported by turbulence and magnetic fields against gravitational collapse. Finally, we identified for the first time a massive ($\sim$200 M$_{\sun}$), collapsing starless clump candidate, "c8", in G035.39-00.33. The magnetic field surrounding "c8" is likely pinched, hinting at an accretion flow along the filament.

astro-ph.GA↗

arXiv2018

Planck Cold Clumps in the $λ$ Orionis complex. II. Environmental effects on core formation

Based on the 850 $μ$m dust continuum data from SCUBA-2 at James Clerk Maxwell Telescope (JCMT), we compare overall properties of Planck Galactic Cold Clumps (PGCCs) in the $λ$ Orionis cloud to those of PGCCs in the Orion A and B clouds. The Orion A and B clouds are well known active star-forming regions, while the $λ$ Orionis cloud has a different environment as a consequence of the interaction with a prominent OB association and a giant Hii region. PGCCs in the $λ$ Orionis cloud have higher dust temperatures ($Td=16.13\pm0.15$ K) and lower values of dust emissivity spectral index ($ β=1.65\pm0.02$) than PGCCs in the Orion A (Td=13.79$\pm 0.21$K, $β=2.07\pm0.03$) and Orion B ($Td=13.82\pm0.19$K, $β=1.96\pm0.02$) clouds. We find 119 sub-structures within the 40 detected PGCCs and identify them as cores. Of total 119 cores, 15 cores are discovered in the $λ$ Orionis cloud, while 74 and 30 cores are found in the Orion A and B clouds, respectively. The cores in the $λ$ Orionis cloud show much lower mean values of size R=0.08 pc, column density N(H2)=$(9.5\pm1.2) \times 10^{22}$ cm$^{-2}$, number density n(H2)=$(2.9 \pm 0.4)\times10^{5}$ cm$^{-3}$, and mass $M_{core}$=$1.0\pm0.3$ M$_{\odot}$ compared to the cores in the Orion A (R=0.11pc, $N(H2)=(2.3\pm0.3) \times 10^{23}$ cm$^{-2}$, n(H2)=$(3.8\pm0.5) \times 10^{5}$cm$^{-3}$, and $M_{core}$=$2.4 \pm 0.3$ M$_{\odot}$) and Orion B (R=0.16pc, N(H2)=$(3.8 \pm 0.4) \times 10^{23}$cm$^{-2}$, n(H2)=$(15.6\pm1.8)\times10^{5}$ cm$^{-3}$, and $M_{core}$= $2.7\pm0.3$ M$_{\odot}$) clouds. These core properties in the $λ$ Orionis cloud can be attributed to the photodissociation and external heating by the nearby Hii region, which may prevent the PGCCs from forming gravitationally bound structures and eventually disperse them. These results support the idea of negative stellar feedback on core formation.

astro-ph.GA↗

arXiv2018

The TOP-SCOPE survey of PGCCs: PMO and SCUBA-2 observations of 64 PGCCs in the 2nd Galactic Quadrant

In order to understand the initial conditions and early evolution of star formation in a wide range of Galactic environments, we carried out an investigation of 64 \textit{Planck} Galactic Cold Clumps (PGCCs) in the second quadrant of the Milky Way. Using the $^{13}$CO and C$^{18}$O $J = 1 - 0$ lines, and 850\,$μ$m continuum observations, we investigated cloud fragmentation and evolution associated with star formation. We extracted 468 clumps and 117 cores from the $^{13}$CO line and 850\,$μ$m continuum maps, respectively. We make use of the Bayesian Distance Calculator and derived the distances of all 64 PGCCs. We found that in general, the mass-size plane follows a relation of $m\sim r^{1.67}$. At a given scale, the masses of our objects are around 1/10 of that of typical Galactic massive star-forming regions. Analysis of the clump and core masses, virial parameters, densities, and mass-size relation suggests that the PGCCs in our sample have a low core formation efficiency ($\sim$3.0\%), and most PGCCs are likely low-mass star-forming candidates. Statistical study indicates that the 850\,$μ$m cores are more turbulent, more optically thick, and denser than the $^{13}$CO clumps for star formation candidates, suggesting that the 850\,$μ$m cores are likely more appropriate future star-formation candidates than the $^{13}$CO clumps.

astro-ph.GA↗

arXiv2018

Top-k Route Search through Submodularity Modeling of Recurrent POI Features

We consider a practical top-k route search problem: given a collection of points of interest (POIs) with rated features and traveling costs between POIs, a user wants to find k routes from a source to a destination and limited in a cost budget, that maximally match her needs on feature preferences. One challenge is dealing with the personalized diversity requirement where users have various trade-off between quantity (the number of POIs with a specified feature) and variety (the coverage of specified features). Another challenge is the large scale of the POI map and the great many alternative routes to search. We model the personalized diversity requirement by the whole class of submodular functions, and present an optimal solution to the top-k route search problem through indices for retrieving relevant POIs in both feature and route spaces and various strategies for pruning the search space using user preferences and constraints. We also present promising heuristic solutions and evaluate all the solutions on real life data.

cs.SI↗

arXiv2018

Status report of the ESCULAP project at Orsay: External injection of low energy electrons in a Plasma

The ESCULAP project aims at studying external injection of low energy (\SI{10}{MeV}) electrons in a plasma in the quasilinear regime. This facility will use the photo injector PHIL and the high power laser LASERIX. We will give a status report of the preliminary work on the facility and the status of the two machines. We will also present the results of simulations showing the expected performances of the facility.

physics.acc-ph↗