arXiv subjects

Kun Xu

Publications and source records attributed to Kun Xu.

At least 253 records · Page 14Linked to original sources

arXiv2021

Two-step multi-resolution reconstruction-based compact gas-kinetic scheme on tetrahedral mesh

In this paper, a third-order compact gas-kinetic scheme (GKS) on unstructured tetrahedral mesh is constructed for the compressible Euler and Navier-Stokes solutions. The time-dependent gas distribution function at a cell interface is used to calculate the fluxes for the updating the cell-averaged flow variables and to evaluate the time accurate cell-averaged flow variables as well for evolving the cell-averaged gradients of flow variables. With the accurate evolution model for both flow variables and their slopes, the quality of the scheme depends closely on the accuracy and reliability of the initial reconstruction of flow variables. The reconstruction scheme becomes more challenge on tetrahedral mesh, where the conventional second-order unlimited least-square reconstruction can make the scheme be linearly unstable when using cell-averaged conservative variables alone with von Neumann neighbors. Benefiting from the evolved cell-averaged slopes, on tetrahedral mesh the GKS is linearly stable from a compact third-order smooth reconstruction with a large CFL number. In order to further increase the robustness of the high-order compact GKS for capturing discontinuous solution, a new two-step multi-resolution weighted essentially non-oscillatory (WENO) reconstruction will be proposed. The novelty of the reconstruction includes the following. Firstly, it releases the stability issue from a second-order compact reconstruction through the introduction of a pre-reconstruction step. Secondly, in the third-order non-linear reconstruction, only one more large stencil is added beside those in the second-order one, which significantly simplifies the high-order reconstruction. The proposed third-order scheme shows good robustness in high speed flow computation and favorable mesh adaptability in cases with complex geometry.

physics.comp-ph↗

arXiv2021

Joint Coreference Resolution and Character Linking for Multiparty Conversation

Character linking, the task of linking mentioned people in conversations to the real world, is crucial for understanding the conversations. For the efficiency of communication, humans often choose to use pronouns (e.g., "she") or normal phrases (e.g., "that girl") rather than named entities (e.g., "Rachel") in the spoken language, which makes linking those mentions to real people a much more challenging than a regular entity linking task. To address this challenge, we propose to incorporate the richer context from the coreference relations among different mentions to help the linking. On the other hand, considering that finding coreference clusters itself is not a trivial task and could benefit from the global character information, we propose to jointly solve these two tasks. Specifically, we propose C$^2$, the joint learning model of Coreference resolution and Character linking. The experimental results demonstrate that C$^2$ can significantly outperform previous works on both tasks. Further analyses are conducted to analyze the contribution of all modules in the proposed model and the effect of all hyper-parameters.

cs.CL↗

arXiv2021

Programmable Multifunctional Plasmonic Waveguide System based on Coding Metamaterials and Inverse Design

In this article, we propose a programmable plasmonic waveguide system (PPWS) to achieve several different functions based on metal coding metamaterials (MCMs) and inverse design technology. There is no need to spend much time on considering the relation between the function and the structure because the MCMs in the PPWS are reprogrammable. In order to demonstrate the effectiveness of the PPWS, we utilize it to achieve several filtering functions, including bandstop and bandpass filters. The simulation results exhibit that the performance of filters is improved based on genetic algorithm, particle swarm optimization, multi-traversal direct-binary search and simulated annealing. Especially, the bandwidth and quality factor for the narrow-band filter can reach 6.5 nm and 200.5. In addition to the simple filtering functions, some relatively complex transmission characteristics can be obtained by using the PPWS, such as plasmon-induced transparency-like effects. In conclusion, genetic algorithm is considered as the most efficient inverse design method for our system due to its less optimization time and stable performance. In comparison with the previous works, our proposed PPWS not only provides a general framework for obtaining an effective, flexible and compact plasmonic device but also shows the applications of inverse design on photonics devices.

physics.optics↗

arXiv2020

TexSmart: A Text Understanding System for Fine-Grained NER and Enhanced Semantic Analysis

This technique report introduces TexSmart, a text understanding system that supports fine-grained named entity recognition (NER) and enhanced semantic analysis functionalities. Compared to most previous publicly available text understanding systems and tools, TexSmart holds some unique features. First, the NER function of TexSmart supports over 1,000 entity types, while most other public tools typically support several to (at most) dozens of entity types. Second, TexSmart introduces new semantic analysis functions like semantic expansion and deep semantic representation, that are absent in most previous systems. Third, a spectrum of algorithms (from very fast algorithms to those that are relatively slow but more accurate) are implemented for one function in TexSmart, to fulfill the requirements of different academic and industrial applications. The adoption of unsupervised or weakly-supervised algorithms is especially emphasized, with the goal of easily updating our models to include fresh data with less human annotation efforts. The main contents of this report include major functions of TexSmart, algorithms for achieving these functions, how to use the TexSmart toolkit and Web APIs, and evaluation results of some key algorithms.

cs.CL↗

arXiv2020

Robust Dialogue Utterance Rewriting as Sequence Tagging

The task of dialogue rewriting aims to reconstruct the latest dialogue utterance by copying the missing content from the dialogue context. Until now, the existing models for this task suffer from the robustness issue, i.e., performances drop dramatically when testing on a different domain. We address this robustness issue by proposing a novel sequence-tagging-based model so that the search space is significantly reduced, yet the core of this task is still well covered. As a common issue of most tagging models for text generation, the model's outputs may lack fluency. To alleviate this issue, we inject the loss signal from BLEU or GPT-2 under a REINFORCE framework. Experiments show huge improvements of our model over the current state-of-the-art systems on domain transfer.

cs.CL↗

arXiv2020

Switching dynamics of reconfigurable perfect soliton crystals in dual-coupled microresonators

Dual-coupled structure is typically used to actively change the local dispersion of microresonator through controllable avoided mode crossings (AMXs). In this paper, we investigate the reconfigurability of perfect soliton crystals (PSCs) based on dual-coupled microresonators. The switching dynamics of PSCs are numerically simulated using perturbed Lugiato-Lefever equation (LLE). Nonlinear phenomena such as solitons rearranging, merging and bursting are observed in the switching process. Specially, for the first time, we have discovered an unexplored $PSC$ $region$ in the microcomb power-detuning phase plane. In $PSC$ $region$, the soliton number ($N$) of PSC state can be switched successively and bidirectionally in a defect-free fashion, verifying the feasibility and advantages of our scheme. The reconfigurability of PSCs would further liberate the application potential of microcombs in a wide range of fields, including frequency metrology, optical communications, and signal-processing systems.

physics.optics↗

arXiv2020

Boosting Adversarial Training with Hypersphere Embedding

Adversarial training (AT) is one of the most effective defenses against adversarial attacks for deep learning models. In this work, we advocate incorporating the hypersphere embedding (HE) mechanism into the AT procedure by regularizing the features onto compact manifolds, which constitutes a lightweight yet effective module to blend in the strength of representation learning. Our extensive analyses reveal that AT and HE are well coupled to benefit the robustness of the adversarially trained models from several aspects. We validate the effectiveness and adaptability of HE by embedding it into the popular AT frameworks including PGD-AT, ALP, and TRADES, as well as the FreeAT and FastAT strategies. In the experiments, we evaluate our methods under a wide range of adversarial attacks on the CIFAR-10 and ImageNet datasets, which verifies that integrating HE can consistently enhance the model robustness for each AT framework with little extra computation.

cs.LG↗

arXiv2020

Efficient Learning of Generative Models via Finite-Difference Score Matching

Several machine learning applications involve the optimization of higher-order derivatives (e.g., gradients of gradients) during training, which can be expensive in respect to memory and computation even with automatic differentiation. As a typical example in generative modeling, score matching (SM) involves the optimization of the trace of a Hessian. To improve computing efficiency, we rewrite the SM objective and its variants in terms of directional derivatives, and present a generic strategy to efficiently approximate any-order directional derivative with finite difference (FD). Our approximation only involves function evaluations, which can be executed in parallel, and no gradient computations. Thus, it reduces the total computational cost while also improving numerical stability. We provide two instantiations by reformulating variants of SM objectives into the FD forms. Empirically, we demonstrate that our methods produce results comparable to the gradient-based counterparts while being much more computationally efficient.

cs.LG↗

arXiv2020

Modeling and computation for non-equilibrium gas dynamics: beyond kinetic relaxation model

The non-equilibrium gas dynamics is described by the Boltzmann equation, which can be solved numerically through the deterministic and stochastic methods. Due to the complicated collision term of the Boltzmann equation, many kinetic relaxation models have been proposed and used in the past seventy years for the study of rarefied flow. In order to develop a multiscale method for the rarefied and continuum flow simulation, by adopting the integral solution of the kinetic model equation a DVM-type unified gas-kinetic scheme (UGKS) has been constructed. The UGKS models the gas dynamics on the cell size and time step scales while the accumulating effect from particle transport and collision has been taken into account within a time step. Under the UGKS framework, a unified gas-kinetic wave-particle (UGKWP) method has been further developed for non-equilibrium flow simulation, where the time evolution of gas distribution function is composed of analytical wave and individual particle. In the highly rarefied regime, particle transport and collision will play a dominant role. Due to the single relaxation time model for particle collision, there is a noticeable discrepancy between the UGKWP solution and the full Boltzmann or DSMC result, especially in the high Mach and Knudsen number cases. In this paper, besides the kinetic relaxation model, a modification of particle collision time according to the particle velocity will be implemented in UGKWP. As a result, the new model greatly improves the performance of UGKWP in the capturing of non-equilibrium flow. There is a perfect match between UGKWP and DSMC or Boltzmann solution in the highly rarefied regime. In the near continuum and continuum flow regime, the UGKWP will gradually get back to the macroscopic variables based Navier-Stokes flow solver at small cell Knudsen number.

physics.comp-ph↗

arXiv2020

A compact high-order gas-kinetic scheme on unstructured mesh for acoustic and shock wave computations

In this paper an even higher-order compact GKS up to sixth order of accuracy will be constructed for the shock and acoustic wave computation on unstructured mesh. The compactness is defined by the physical domain of dependence for an unstructured triangular cell, which may involve the closest neighbors of neighboring cells. The compactness and high-order accuracy of the scheme are coming from the consistency between the high-order initial reconstruction and the high-order gas evolution model under GKS framework. The high-order evolution solution at a cell interface provides not only a time-accurate flux function, but also the time-evolving flow variables. Therefore, the cell-averaged flow variables and their gradients can be explicitly updated at the next time level from the moments of the same time-dependent gas distribution function. Based on the cell averages and cell-averaged derivatives, both linear and nonlinear high-order reconstruction can be obtained for macroscopic flow variables in the evaluation of local equilibrium and non-equilibrium states. The current nonlinear reconstruction is a combination of WENO and ENO methodology. The initial piecewise discontinuous reconstruction is used for the determination non-equilibrium state and an evolved smooth reconstruction for the equilibrium state. The evolution model in gas-kinetic scheme is based on a relaxation process from non-equilibrium to equilibrium state. The accuracy, efficiency, and robustness of the scheme have been validated. The main conclusion of the paper is that beyond the first-order Riemann solver, the use of high-order gas evolution model seems necessary in the development of high-order schemes.

math.NA↗

arXiv2020

Bi-level Score Matching for Learning Energy-based Latent Variable Models

Score matching (SM) provides a compelling approach to learn energy-based models (EBMs) by avoiding the calculation of partition function. However, it remains largely open to learn energy-based latent variable models (EBLVMs), except some special cases. This paper presents a bi-level score matching (BiSM) method to learn EBLVMs with general structures by reformulating SM as a bi-level optimization problem. The higher level introduces a variational posterior of the latent variables and optimizes a modified SM objective, and the lower level optimizes the variational posterior to fit the true posterior. To solve BiSM efficiently, we develop a stochastic optimization algorithm with gradient unrolling. Theoretically, we analyze the consistency of BiSM and the convergence of the stochastic algorithm. Empirically, we show the promise of BiSM in Gaussian restricted Boltzmann machines and highly nonstructural EBLVMs parameterized by deep convolutional neural networks. BiSM is comparable to the widely adopted contrastive divergence and SM methods when they are applicable; and can learn complex EBLVMs with intractable posteriors to generate natural images.

cs.LG↗

arXiv2020

Unified gas-kinetic wave-particle methods V: diatomic molecular flow

In this paper, the unified gas-kinetic wave-particle (UGKWP) method is further developed for diatomic gas with the energy exchange between translational and rotational modes for flow study in all regimes. The multiscale transport mechanism in UGKWP is coming from the direct modeling in a discretized space, where the cell's Knudsen number, defined by the ratio of particle mean free path over the numerical cell size, determines the flow physics simulated by the wave particle formulation. The non-equilibrium distribution function in UGKWP is tracked by the discrete particle and analytical wave. The weights of distributed particle and wave in different regimes are controlled by the accumulating evolution solution of particle transport and collision within a time step, where distinguishable macroscopic flow variables of particle and wave are updated inside each control volume. With the variation of local cell's Knudsen number, the UGKWP becomes a particle method in the highly rarefied flow regime and converges to the gas-kinetic scheme (GKS) for the Navier-Stokes solution in the continuum flow regime without particles. Even targeting on the same solution as the discrete velocity method (DVM)-based unified gas-kinetic scheme (UGKS), the computational cost and memory requirement in UGKWP could be reduced by several orders of magnitude for the high speed and high temperature flow simulation, where the translational and rotational non-equilibrium becomes important in the transition and rarefied regime. As a result, 3D hypersonic computations around a flying vehicle in all regimes can be conducted using a personal computer. The UGKWP method for diatomic gas will be validated in various cases from one dimensional shock structure to three dimensional flow over a sphere, and the numerical solutions will be compared with the reference DSMC results and experimental measurements.

physics.comp-ph↗

arXiv2020

Semantic Role Labeling Guided Multi-turn Dialogue ReWriter

For multi-turn dialogue rewriting, the capacity of effectively modeling the linguistic knowledge in dialog context and getting rid of the noises is essential to improve its performance. Existing attentive models attend to all words without prior focus, which results in inaccurate concentration on some dispensable words. In this paper, we propose to use semantic role labeling (SRL), which highlights the core semantic information of who did what to whom, to provide additional guidance for the rewriter model. Experiments show that this information significantly improves a RoBERTa-based model that already outperforms previous state-of-the-art systems.

cs.CL↗

arXiv2020

Microwave Photonic Ising Machine

Ising machines based on analog systems have the potential of acceleration in solving ubiquitous combinatorial optimization problems. Although some artificial spins to support large-scale Ising machine is reported, e.g. superconducting qubits in quantum annealers and short optical pulses in coherent Ising machines, the spin coherence is fragile due to the ultra-low equivalent temperature or optical phase sensitivity. In this paper, we propose to use short microwave pulses generated from an optoelectronic parametric oscillator as the spins to implement the Ising machine with large scale and also high coherence under room temperature. The proposed machine supports 10,000 spins, and the high coherence leads to accurate computation. Moreover, the Ising machine is highly compatible with high-speed electronic devices for programmability, paving a low-cost, accurate, and easy-to-implement way toward to solve real-world optimization problems.

physics.app-ph↗

arXiv2020

Unified Gas-kinetic Wave-Particle Method IV: Multi-species Gas Mixture and Plasma Transport

In this paper, we extend the unified gas-kinetic wave-particle (UGKWP) method to the multi-species gas mixture and multiscale plasma transport. The construction of the scheme is based on the direct modeling on the mesh size and time step scales, and the local cell's Knudsen number determines the flow physics. The proposed scheme has the multiscale and asymptotic complexity diminishing properties. The multiscale property means that according to cell's Knudsen number the scheme can capture the non-equilibrium flow physics in the rarefied flow regime, and preserve the asymptotic Euler, Navier-Stokes, and magnetohydrodynamics limit in the continuum regime. The asymptotic complexity diminishing property means that the total degree of freedom of the scheme automatically decreases as cell's Knudsen number decreases. In the continuum regime, the scheme automatically degenerates from a kinetic solver to a hydrodynamic solver. In UGKWP, the evolution of microscopic velocity distribution is coupled with the evolution of macroscopic variables, and the particle evolution as well as the macroscopic fluxes are modeled from the time accumulating solution up to a time step scale from the kinetic model equation. For plasma transport, current scheme provides a smooth transition from particle in cell (PIC) method in the rarefied regime to the magnetohydrodynamic (MHD) solver in the continuum regime. In the continuum limit, the cell size and time step of the UGKWP method is not restricted to be less than the mean free path and mean collision time. In the highly magnetized regime, the cell size and time step are not restricted by the Debye length and plasma cyclotron period. The multiscale and asymptotic complexity diminishing properties of the scheme are verified by numerical tests in multiple flow regimes.

physics.comp-ph↗

arXiv2020

Triple Generative Adversarial Networks

We propose a unified game-theoretical framework to perform classification and conditional image generation given limited supervision. It is formulated as a three-player minimax game consisting of a generator, a classifier and a discriminator, and therefore is referred to as Triple Generative Adversarial Network (Triple-GAN). The generator and the classifier characterize the conditional distributions between images and labels to perform conditional generation and classification, respectively. The discriminator solely focuses on identifying fake image-label pairs. Under a nonparametric assumption, we prove the unique equilibrium of the game is that the distributions characterized by the generator and the classifier converge to the data distribution. As a byproduct of the three-player mechanism, Triple-GAN is flexible to incorporate different semi-supervised classifiers and GAN architectures. We evaluate Triple-GAN in two challenging settings, namely, semi-supervised learning and the extreme low data regime. In both settings, Triple-GAN can achieve excellent classification results and generate meaningful samples in a specific class simultaneously. In particular, using a commonly adopted 13-layer CNN classifier, Triple-GAN outperforms extensive semi-supervised learning methods substantially on more than 10 benchmarks no matter data augmentation is applied or not.

cs.LG↗

arXiv2020

Comparison of the performance of high-order schemes based on the gas-kinetic and HLLC fluxes

In this paper, a comparison of the performance of two high-order finite volume methods based on the gas-kinetic scheme (GKS) and HLLC fluxes is carried out in structured rectangular mesh. For both schemes, the fifth-order WENO-AO reconstruction is adopted to achieve a high-order spatial accuracy. In terms of temporal discretization, a two-stage fourth-order (S2O4) time marching strategy is adopted for WENO5-AO-GKS scheme, and the fourth-order Runge-Kutta (RK4) method is employed for WENO5-AO-HLLC scheme. For the viscous flow computation, the GKS includes both inviscid and viscous fluxes in the evolution of a single cell interface gas distribution function. While for the WENO5-AO-HLLC scheme, the inviscid flux is provided by HLLC Riemann solver, and the viscous flux is discretized by a sixth-order central difference method. Based on the tests of forward Mach step and viscous shock tube, both schemes show outstanding shock capturing property. From the Titarev-Toro and double shear layer tests, WENO5-AO-GKS scheme seems to have a better resolution than WENO5-AO-HLLC scheme. Both schemes show excellent robustness in extreme cases, such as the Le Blanc problem. From the cases of the Noh problem and the compressible isotropic turbulence, WENO5-AO-GKS scheme shows favorite robustness. In the compressible isotropic turbulence and three-dimensional Taylor-Green vortex problems, WENO-AO-GKS can use a CFL number up to 0.5, instead of 0.3 for WENO5-AO-HLLC. In terms of computational efficiency, WENO5-AO-HLLC scheme is about 27% more expensive than WENO5-AO-GKS scheme in the two-dimensional viscous flow problems, but is about 15% faster in the three-dimensional case. Due to the multi-dimensionality, WENO5-AO-GKS scheme performs better than WENO5-AO-HLLC scheme in the laminar boundary layer and the double shear layer test.

physics.comp-ph↗

arXiv2020

A three-dimensional compact high-order gas-kinetic scheme on structured mesh

In this paper, a third-order compact gas-kinetic scheme is firstly proposed for three-dimensional computation for the compressible Euler and Navier-Stokes solutions. The scheme achieves its compactness due to the time-dependent gas distribution function in GKS, which provides not only the fluxes but also the time accurate flow variables in the next time level at a cell interface. As a result, the cell averaged first-order spatial derivatives of flow variables can be obtained naturally through the Gauss's theorem. Then, a third-order compact reconstruction involving the cell averaged values and their first-order spatial derivatives can be achieved. The trilinear interpolation is used to treat possible non-coplanar elements on general hexahedral mesh. The constrained least-square technique is applied to improve the accuracy in the smooth case. To deal with both smooth and discontinuous flows, a new HWENO reconstruction is designed in the current scheme by following the ideas in Zhu, 2018. No identification of troubled cells is needed in the current scheme. In contrast to the Riemann solver-based method, the compact scheme can achieve a third-order temporal accuracy with the two-stage two-derivative temporal discretization, instead of the three-stage Runge-Kutta method. Overall, the proposed scheme inherits the high accuracy and efficiency of the previous ones in two-dimensional case. The desired third-order accuracy can be obtained with curved boundary. The robustness of the scheme has been validated through many cases, including strong shocks in both inviscid and viscous flow computations. Quantitative comparisons for both smooth and discontinuous cases show that the current third-order scheme can give competitive results against the fifth-order non-compact GKS under the same mesh. A large CFL number around 0.5 can be used in the present scheme.

physics.comp-ph↗