arXiv subjects

Cheng Wang

Publications and source records attributed to Cheng Wang.

At least 379 records · Page 21Linked to original sources

arXiv2021

Learning Slice-Aware Representations with Mixture of Attentions

Real-world machine learning systems are achieving remarkable performance in terms of coarse-grained metrics like overall accuracy and F-1 score. However, model improvement and development often require fine-grained modeling on individual data subsets or slices, for instance, the data slices where the models have unsatisfactory results. In practice, it gives tangible values for developing such models that can pay extra attention to critical or interested slices while retaining the original overall performance. This work extends the recent slice-based learning (SBL)~\cite{chen2019slice} with a mixture of attentions (MoA) to learn slice-aware dual attentive representations. We empirically show that the MoA approach outperforms the baseline method as well as the original SBL approach on monitored slices with two natural language understanding (NLU) tasks.

cs.LG↗

arXiv2021

DLA-Net: Learning Dual Local Attention Features for Semantic Segmentation of Large-Scale Building Facade Point Clouds

Semantic segmentation of building facade is significant in various applications, such as urban building reconstruction and damage assessment. As there is a lack of 3D point clouds datasets related to the fine-grained building facade, we construct the first large-scale building facade point clouds benchmark dataset for semantic segmentation. The existing methods of semantic segmentation cannot fully mine the local neighborhood information of point clouds. Addressing this problem, we propose a learnable attention module that learns Dual Local Attention features, called DLA in this paper. The proposed DLA module consists of two blocks, including the self-attention block and attentive pooling block, which both embed an enhanced position encoding block. The DLA module could be easily embedded into various network architectures for point cloud segmentation, naturally resulting in a new 3D semantic segmentation network with an encoder-decoder architecture, called DLA-Net in this work. Extensive experimental results on our constructed building facade dataset demonstrate that the proposed DLA-Net achieves better performance than the state-of-the-art methods for semantic segmentation.

cs.CV↗

arXiv2021

Convergence analysis of the variational operator splitting scheme for a reaction-diffusion system with detailed balance

We present a detailed convergence analysis for an operator splitting scheme proposed in [C. Liu et al.,J. Comput. Phys., 436, 110253, 2021] for a reaction-diffusion system with detailed balance. The numerical scheme has been constructed based on a recently developed energetic variational formulation, in which the reaction part is reformulated in terms of the reaction trajectory, and both the reaction and diffusion parts dissipate the same free energy. The scheme is energy stable and positivity-preserving. In this paper, the detailed convergence analysis and error estimate are performed for the operator splitting scheme. The nonlinearity in the reaction trajectory equation, as well as the implicit treatment of nonlinear and singular logarithmic terms, impose challenges in numerical analysis. To overcome these difficulties, we make use of the convex nature of the logarithmic nonlinear terms, which are treated implicitly in the chemical reaction stage. In addition, a combination of a rough error estimate and a refined error estimate leads to a desired bound of the numerical error in the reaction stage, in the discrete maximum norm. Furthermore, a discrete maximum principle yields the evolution bound of the numerical error function at the diffusion stage. As a direct consequence, a combination of the numerical error analysis at different stages and the consistency estimate for the operator splitting results in the convergence estimate of the numerical scheme for the full reaction-diffusion system.

math.NA↗

arXiv2021

A second-order numerical method for Landau-Lifshitz-Gilbert equation with large damping parameters

A second order accurate numerical scheme is proposed and implemented for the Landau-Lifshitz-Gilbert equation, which models magnetization dynamics in ferromagnetic materials, with large damping parameters. The main advantages of this method are associated with the following features: (1) It only solves linear systems of equations with constant coefficients where fast solvers are available, so that the numerical efficiency has been greatly improved, in comparison with the existing Gauss-Seidel project method. (2) The second-order accuracy in time is achieved, and it is unconditionally stable for large damping parameters. Moreover, both the second-order accuracy and the great efficiency improvement will be verified by several numerical examples in the 1D and 3D simulations. In the presence of large damping parameters, it is observed that this method is unconditionally stable and finds physically reasonable structures while many existing methods have failed. For the domain wall dynamics, the linear dependence of wall velocity with respect to the damping parameter and the external magnetic field will be obtained through the reported simulations.

physics.comp-ph↗

arXiv2021

A positivity-preserving, energy stable scheme for a Ternary Cahn-Hilliard system with the singular interfacial parameters

In this paper, we construct and analyze a uniquely solvable, positivity preserving and unconditionally energy stable finite-difference scheme for the periodic three-component Macromolecular Microsphere Composite (MMC) hydrogels system, a ternary Cahn-Hilliard system with a Flory-Huggins-deGennes free energy potential. The proposed scheme is based on a convex-concave decomposition of the given energy functional with two variables, and the centered difference method is adopted in space. We provide a theoretical justification that this numerical scheme has a pair of unique solutions, such that the positivity is always preserved for all the singular terms, i.e., not only two phase variables are always between $0$ and $1$, but also the sum of two phase variables is between $0$ and $1$, at a point-wise level. In addition, we use the local Newton approximation and multigrid method to solve this nonlinear numerical scheme, and various numerical results are presented, including the numerical convergence test, positivity-preserving property test, energy dissipation and mass conservation properties.

math.NA↗

arXiv2021

Lagrangian dynamics and heat transfer in porous media convection

We report a numerical study of Rayleigh--Bénard convection through random porous media using pore-scale modelling, focusing on the Lagrangian dynamics of fluid particles and heat transfer for varied porosities $ϕ$. Due to the interaction between the porous medium and the coherent flow structures, the flow is found to be highly heterogeneous, consisting of convection channels with strong flow strength and stagnant regions with low velocities. The modifications of flow field due to porous structure have a significant influence on the dynamics of fluid particles. Evaluation of the particle displacement along the trajectory reveals the emergence of anomalous transport for long times as $ϕ$ is decreased, which is associated with the long-time correlation of Lagrangian velocity of the fluid. As porosity is decreased, the cross-correlation between the vertical velocity and temperature fluctuation is enhanced, which reveals a mechanism to enhance the heat transfer in porous media convection.

physics.flu-dyn↗

arXiv2021

Handling Long-Tail Queries with Slice-Aware Conversational Systems

We have been witnessing the usefulness of conversational AI systems such as Siri and Alexa, directly impacting our daily lives. These systems normally rely on machine learning models evolving over time to provide quality user experience. However, the development and improvement of the models are challenging because they need to support both high (head) and low (tail) usage scenarios, requiring fine-grained modeling strategies for specific data subsets or slices. In this paper, we explore the recent concept of slice-based learning (SBL) (Chen et al., 2019) to improve our baseline conversational skill routing system on the tail yet critical query traffic. We first define a set of labeling functions to generate weak supervision data for the tail intents. We then extend the baseline model towards a slice-aware architecture, which monitors and improves the model performance on the selected tail intents. Applied to de-identified live traffic from a commercial conversational AI system, our experiments show that the slice-aware model is beneficial in improving model performance for the tail intents while maintaining the overall performance.

cs.LG↗

arXiv2021

An energy stable finite element scheme for the three-component Cahn-Hilliard-type model for macromolecular microsphere composite hydrogels

In this article, we present and analyze a finite element numerical scheme for a three-component macromolecular microsphere composite (MMC) hydrogel model, which takes the form of a ternary Cahn-Hilliard-type equation with Flory-Huggins-deGennes energy potential. The numerical approach is based on a convex-concave decomposition of the energy functional in multi-phase space, in which the logarithmic and the nonlinear surface diffusion terms are treated implicitly, while the concave expansive linear terms are explicitly updated. A mass lumped finite element spatial approximation is applied, to ensure the positivity of the phase variables. In turn, a positivity-preserving property can be theoretically justified for the proposed fully discrete numerical scheme. In addition, unconditional energy stability is established as well, which comes from the convexity analysis. Several numerical simulations are carried out to verify the accuracy and positivity-preserving property of the proposed scheme.

math.NA↗

arXiv2021

FAIR1M: A Benchmark Dataset for Fine-grained Object Recognition in High-Resolution Remote Sensing Imagery

With the rapid development of deep learning, many deep learning-based approaches have made great achievements in object detection task. It is generally known that deep learning is a data-driven method. Data directly impact the performance of object detectors to some extent. Although existing datasets have included common objects in remote sensing images, they still have some limitations in terms of scale, categories, and images. Therefore, there is a strong requirement for establishing a large-scale benchmark on object detection in high-resolution remote sensing images. In this paper, we propose a novel benchmark dataset with more than 1 million instances and more than 15,000 images for Fine-grAined object recognItion in high-Resolution remote sensing imagery which is named as FAIR1M. All objects in the FAIR1M dataset are annotated with respect to 5 categories and 37 sub-categories by oriented bounding boxes. Compared with existing detection datasets dedicated to object detection, the FAIR1M dataset has 4 particular characteristics: (1) it is much larger than other existing object detection datasets both in terms of the quantity of instances and the quantity of images, (2) it provides more rich fine-grained category information for objects in remote sensing images, (3) it contains geographic information such as latitude, longitude and resolution, (4) it provides better image quality owing to a careful data cleaning procedure. To establish a baseline for fine-grained object recognition, we propose a novel evaluation method and benchmark fine-grained object detection tasks and a visual classification task using several State-Of-The-Art (SOTA) deep learning-based models on our FAIR1M dataset. Experimental results strongly indicate that the FAIR1M dataset is closer to practical application and it is considerably more challenging than existing datasets.

cs.CV↗

arXiv2021

Neighbor Embedding Variational Autoencoder

Being one of the most popular generative framework, variational autoencoders(VAE) are known to suffer from a phenomenon termed posterior collapse, i.e. the latent variational distributions collapse to the prior, especially when a strong decoder network is used. In this work, we analyze the latent representation of collapsed VAEs, and proposed a novel model, neighbor embedding VAE(NE-VAE), which explicitly constraints the encoder to encode inputs close in the input space to be close in the latent space. We observed that for VAE variants that report similar ELBO, KL divergence or even mutual information scores may still behave quite differently in the latent organization. In our experiments, NE-VAE can produce qualitatively different latent representations with majority of the latent dimensions remained active, which may benefit downstream latent space optimization tasks. NE-VAE can prevent posterior collapse to a much greater extent than it's predecessors, and can be easily plugged into any autoencoder framework, without introducing addition model components and complex training routines.

cs.LG↗

arXiv2021

A structure-preserving, operator splitting scheme for reaction-diffusion equations with detailed balance

In this paper, we propose and analyze a positivity-preserving, energy stable numerical scheme for certain type reaction-diffusion systems involving the Law of Mass Action with the detailed balance condition. The numerical scheme is constructed based on a recently developed energetic variational formulation, in which the reaction part is reformulated in terms of reaction trajectories. The fact that both the reaction and the diffusion parts dissipate the same free energy opens a path of an energy stable, operator splitting scheme for these systems. At the reaction stage, we solve equations of reaction trajectories by treating all the logarithmic terms in the reformulated form implicitly due to their convex nature. The positivity-preserving property and unique solvability can be theoretically proved, based on the singular behavior of the logarithmic function around the limiting value. Moreover, the energy stability of this scheme at the reaction stage can be proved by a careful convexity analysis. Similar techniques are used to establish the positivity-preserving property and energy stability for the standard semi-implicit solver at the diffusion stage. As a result, a combination of these two stages leads to a positivity-preserving and energy-stable numerical scheme for the original reaction-diffusion system. To our best knowledge, it is the first time to report an energy-dissipation-law-based operator splitting scheme to a nonlinear PDE with variational structures. Several numerical examples are presented to demonstrate the robustness of the proposed operator splitting scheme.

math.NA↗

arXiv2021

A positivity-preserving and convergent numerical scheme for the binary fluid-surfactant system

In this paper, we develop a first order (in time) numerical scheme for the binary fluid surfactant phase field model. The free energy contains a double-well potential, a nonlinear coupling entropy and a Flory-Huggins potential. The resulting coupled system consists of two Cahn-Hilliard type equations. This system is solved numerically by finite difference spatial approximation, in combination with convex splitting temporal discretization. We prove the proposed scheme is unique solvable, positivity-preserving and unconditionally energy stable. In addition, an optimal rate convergence analysis is provided for the proposed numerical scheme, which will be the first such result for the binary fluid-surfactant system. Newton iteration is used to solve the discrete system. Some numerical experiments are performed to validate the accuracy and energy stability of the proposed scheme.

math.NA↗

arXiv2021

VPC-Net: Completion of 3D Vehicles from MLS Point Clouds

As a dynamic and essential component in the road environment of urban scenarios, vehicles are the most popular investigation targets. To monitor their behavior and extract their geometric characteristics, an accurate and instant measurement of vehicles plays a vital role in traffic and transportation fields. Point clouds acquired from the mobile laser scanning (MLS) system deliver 3D information of road scenes with unprecedented detail. They have proven to be an adequate data source in the fields of intelligent transportation and autonomous driving, especially for extracting vehicles. However, acquired 3D point clouds of vehicles from MLS systems are inevitably incomplete due to object occlusion or self-occlusion. To tackle this problem, we proposed a neural network to synthesize complete, dense, and uniform point clouds for vehicles from MLS data, named Vehicle Points Completion-Net (VPC-Net). In this network, we introduce a new encoder module to extract global features from the input instance, consisting of a spatial transformer network and point feature enhancement layer. Moreover, a new refiner module is also presented to preserve the vehicle details from inputs and refine the complete outputs with fine-grained information. Given sparse and partial point clouds as inputs, the network can generate complete and realistic vehicle structures and keep the fine-grained details from the partial inputs. We evaluated the proposed VPC-Net in different experiments using synthetic and real-scan datasets and applied the results to 3D vehicle monitoring tasks. Quantitative and qualitative experiments demonstrate the promising performance of the proposed VPC-Net and show state-of-the-art results.

cs.CV↗

arXiv2021

Efficient erbium-doped thin-film lithium niobate waveguide amplifiers

Lithium niobate on insulator (LNOI) is an emerging photonic platform with great promises for future optical communications, nonlinear optics and microwave photonics. An important integrated photonic building block, active waveguide amplifiers, however, is still missing in the LNOI platform. Here we report an efficient and compact waveguide amplifier based on erbium-doped LNOI waveguides, realized by a sequence of erbium-doped crystal growth, ion slicing and lithography-based waveguide fabrication. Using a compact 5-mm-long waveguide, we demonstrate on-chip net gain of > 5 dB for 1530-nm signal light with a relatively low pump power of 21 mW at 980 nm. The efficient LNOI waveguide amplifiers could become an important fundamental element in future lithium niobate photonic integrated circuits.

physics.optics↗

arXiv2020

A second order accurate scalar auxiliary variable (SAV) numerical method for the square phase field crystal equation

In this paper we propose and analyze a second order accurate (in time) numerical scheme for the square phase field crystal (SPFC) equation, a gradient flow modeling crystal dynamics at the atomic scale in space but on diffusive scales in time. Its primary difference with the standard phase field crystal model is an introduction of the 4-Laplacian term in the free energy potential, which in turn leads to a much higher degree of nonlinearity. To make the numerical scheme linear while preserving the nonlinear energy stability, we make use of the scalar auxiliary variable (SAV) approach, in which a second order backward differentiation formula (BDF) is applied in the temporal stencil. Meanwhile, a direct application of the SAV method faces certain difficulties, due to the involvement of the 4-Laplacian term, combined with a derivation of the lower bound of the nonlinear energy functional. In the proposed numerical method, an appropriate decomposition for the physical energy functional is formulated, so that the nonlinear energy part has a well-established global lower bound, and the rest terms lead to constant-coefficient diffusion terms with positive eigenvalues. In turn, the numerical scheme could be very efficiently implemented by constant-coefficient Poisson-like type solvers (via FFT), and energy stability is established by introducing an auxiliary variable, and an optimal rate convergence analysis is provided for the proposed SAV method. A few numerical experiments are also presented, which confirm the efficiency and accuracy of the proposed scheme.

math.NA↗

arXiv2020

Structure-preserving, energy stable numerical schemes for a liquid thin film coarsening model

In this paper, two finite difference numerical schemes are proposed and analyzed for the droplet liquid film model, with a singular Leonard-Jones energy potential involved. Both first and second order accurate temporal algorithms are considered. In the first order scheme, the convex potential and the surface diffusion terms are implicitly, while the concave potential term is updated explicitly. Furthermore, we provide a theoretical justification that this numerical algorithm has a unique solution, such that the positivity is always preserved for the phase variable at a point-wise level, so that a singularity is avoided in the scheme. In fact, the singular nature of the Leonard-Jones potential term around the value of 0 prevents the numerical solution reaching such singular value, so that the positivity structure is always preserved. Moreover, an unconditional energy stability of the numerical scheme is derived, without any restriction for the time step size. In the second order numerical scheme, the BDF temporal stencil is applied, and an alternate convex-concave decomposition is derived, so that the concave part corresponds to a quadratic energy. In turn, the combined Leonard-Jones potential term is treated implicitly, and the concave part the is approximated by a second order Adams-Bashforth explicit extrapolation, and an artificial Douglas-Dupont regularization term is added to ensure the energy stability. The unique solvability and the positivity-preserving property for the second order scheme could be similarly established. In addition, optimal rate convergence analysis is provided for both the first and second order accurate schemes. A few numerical simulation results are also presented, which demonstrate the robustness of the numerical schemes.

math.NA↗

arXiv2020

Efficient and DoS-resistant Consensus for Permissioned Blockchains

Existing permissioned blockchain systems designate a fixed and explicit group of committee nodes to run a consensus protocol that confirms the same sequence of blocks among all nodes. Unfortunately, when such a permissioned blockchain runs in a large scale on the Internet, these explicit committee nodes can be easily turned down by denial-of-service (DoS) or network partition attacks. Although work proposes scalable BFT protocols that run on a larger number of committee nodes, their efficiency drops dramatically when only a small number of nodes are attacked. In this paper, our EGES protocol leverages Intel SGX to develop a new abstraction called "stealth committee", which effectively hides the committee nodes into a large pool of fake committee nodes. EGES selects a distinct group of stealth committee for each block and confirms the same sequence of blocks among all nodes with overwhelming probability. Evaluation on typical geo-distributed settings shows that: (1)EGES is the first permissioned blockchain's consensus protocol that can tolerate tough DoS and network partition attacks; and (2) EGES achieves comparable throughput and latency as existing permissioned blockchains' protocols

cs.DC↗

arXiv2020

Rotational dynamics of bottom-heavy rods in turbulence from experiments and numerical simulations

We successfully perform the three-dimensional tracking in a turbulent fluid flow of small asymmetrical particles that are neutrally-buoyant and bottom-heavy, i.e., they have a non-homogeneous mass distribution along their symmetry axis. We experimentally show how a tiny mass inhomogeneity can affect the particle orientation along the preferred vertical direction and modify its tumbling rate. The experiment is complemented by a series of simulations based on realistic Navier-Stokes turbulence and on a point-like particle model that is capable to explore the full range of parameter space characterized by the gravitational torque stability number and by the particle aspect ratio. We propose a theoretical perturbative prediction valid in the high bottom-heaviness regime that agrees well with the observed preferential orientation and tumbling rate of the particles. We also show that the heavy-tail shape of the probability distribution function of the tumbling rate is weakly affected by the bottom-heaviness of the particles.

physics.flu-dyn↗