arXiv subjects

Xiaohui Wang

Publications and source records attributed to Xiaohui Wang.

At least 37 records · Page 2Linked to original sources

arXiv2023

Magnetically and electrically controllable valley splittings in MXene monolayers

The modulation of the valley structure in two-dimensional valley materials is vital in the field of valleytronics. The multiferroicity provides possibility for multiple modulations of the valley, including the magnetic and electric means. Based on the first-principle calculations, we study the valley properties and associated manipulations of multiferroic Co$_2$CF$_2$ monolayers with different stacking patterns. Our calculations show that the Co$_2$CF$_2$ monolayer in the H$^{\prime}$ phase is a ferrovalley material, with sizable valley splittings. By rotating the magnetization direction, the valley splittings can be tuned for both the magnitude and sign. The electric field, driving the reversal of the electric polarization, can also change the magnitude of the valley splittings. Besides, a metastable T$^{\prime}$ phase exhibits valley splittings as well, of which the magnitude and sign can be simultaneously controlled by applied magnetic and electric fields. These findings offer a practical way for realizing highly tunable valleys by multiferroic couplings.

cond-mat.mtrl-sci↗

arXiv2023

Understanding Parameter Sharing in Transformers

Parameter sharing has proven to be a parameter-efficient approach. Previous work on Transformers has focused on sharing parameters in different layers, which can improve the performance of models with limited parameters by increasing model depth. In this paper, we study why this approach works from two perspectives. First, increasing model depth makes the model more complex, and we hypothesize that the reason is related to model complexity (referring to FLOPs). Secondly, since each shared parameter will participate in the network computation several times in forward propagation, its corresponding gradient will have a different range of values from the original model, which will affect the model convergence. Based on this, we hypothesize that training convergence may also be one of the reasons. Through further analysis, we show that the success of this approach can be largely attributed to better convergence, with only a small part due to the increased model complexity. Inspired by this, we tune the training hyperparameters related to model convergence in a targeted manner. Experiments on 8 machine translation tasks show that our model achieves competitive performance with only half the model complexity of parameter sharing models.

cs.LG↗

arXiv2023

MobileNMT: Enabling Translation in 15MB and 30ms

Deploying NMT models on mobile devices is essential for privacy, low latency, and offline scenarios. For high model capacity, NMT models are rather large. Running these models on devices is challenging with limited storage, memory, computation, and power consumption. Existing work either only focuses on a single metric such as FLOPs or general engine which is not good at auto-regressive decoding. In this paper, we present MobileNMT, a system that can translate in 15MB and 30ms on devices. We propose a series of principles for model compression when combined with quantization. Further, we implement an engine that is friendly to INT8 and decoding. With the co-design of model and engine, compared with the existing system, we speed up 47.0x and save 99.5% of memory with only 11.6% loss of BLEU. The code is publicly available at https://github.com/zjersey/Lightseq-ARM.

cs.AI↗

arXiv2023

Electron-irradiation-facilitated production of chemically homogenized nanotwins in nanolaminated carbides

Twin boundaries have been exploited to stabilize ultrafine grains and improve the mechanical properties of nanomaterials. The production of twin boundaries and nanotwins is however prohibitively challenging in carbide ceramics. Using scanning transmission electron microscopes as a unique platform for atomic-scale structure engineering, we demonstrate that twin platelets could be produced in carbides by engineering antisite defects. Antisite defects at metal sites in various layered ternary carbides are collectively and controllably generated and the metal elements are homogenized by electron irradiation, which transforms the twin-like lamellae into nanotwin platelets. Accompanying the chemical homogenization, α-Ti3AlC2 transforms to unconventional \b{eta}-Ti3AlC2. The chemical homogeneity and the width of the twin platelets can be tuned by the dose and energy of bombarding electrons. Chemically homogenized nanotwins can boost the hardness by ~45%. Our results provide a new way to produce ultrathin (<5 nm) nanotwin platelets in scientifically and technologically important carbide materials and showcase the feasibility of defect engineering by an angstrom-sized electron probe.

cond-mat.mtrl-sci↗

arXiv2022

Valley and spin splittings in a valley-layer-locked monolayer via atomic adsorption

To date, a number of valley materials have been discovered with the spin-valley or valley-layer couplings. It is highly desirable to realize the interplay of more electronic degrees of freedom in a valley material. Based on the first-principles calculations, we demonstrate the valley and spin degeneracy liftings in the band structure of a TiSiCO monolayer, with the help of the selective expression of the layer pseudospin via atomic adsorption. The introduction of the transition-metal adatoms provides an effective electric field and magnetic proximity effect, giving rise to the valley and spin splittings. These splittings can be further tuned by applying an external electric field. According to the modified band structure, various interlayer excitons with different combinations of spins and electric dipoles are selectively created, under the optical field of appropriate frequencies. The tunable spin and valley splittings in atom adsorbed TiSiCO monolayer offer opportunities for exploring the interactions between spin, valley and layer pseudospin, and designing advanced optoelectronic devices.

cond-mat.mtrl-sci↗

arXiv2022

Tunable valley and spin splittings in $\rm VSi_2N_4$ bilayer

The control and manipulation of the valley and spin degrees of freedom have received great interests in fundamental studies and advanced information technologies. Compared with magnetic means, it is highly desirable to realize more energy-efficient electric control of valley and spin. Using the first-principles calculations, we demonstrate tunable valley and spin degeneracy splittings in $\rm VSi_2N_4$ bilayers, with the aid of the layered structure and associated electric control. Depending on different interlayer magnetic couplings and stacking orders, the $\rm VSi_2N_4$ bilayers exhibit a variety of combinations of valley and spin degeneracies. Under the action of a vertical electric field, the degeneracy splittings become highly tunable for both the sign and magnitude. As a result, a series of anomalous Hall currents can be selectively realized with varied indices of valley and spin. These intriguing features offer a practical way for designing energy-efficient devices based on the couplings between multiple electronic degrees of freedom.

cond-mat.mtrl-sci↗

arXiv2022

LightSeq2: Accelerated Training for Transformer-based Models on GPUs

Transformer-based neural models are used in many AI applications. Training these models is expensive, as it takes huge GPU resources and long duration. It is challenging because typical data like sentences have variable lengths, and Transformer's computation patterns are more complex than convolutional neural networks. Existing systems either only focus on model inference or optimization for only BERT-like encoder models. In this paper, we present LightSeq2, a system to accelerate training for a general family of Transformer models on GPUs. We propose a series of GPU optimization techniques tailored to the specific computation flow and memory access patterns of Transformer models. LightSeq2 supports many model architectures, including BERT (encoder-only), GPT (decoder-only), Transformer (encoder-decoder), and vision Transformer. Our experiments for a variety of models and benchmarks show that LightSeq2 is consistently faster (1.4-3.5x) than previous systems on different GPUs. In particular, it gains 308% training speedup compared with existing systems on a large public machine translation benchmark (WMT14 English-German).

cs.CL↗

arXiv2021

LightSeq: A High Performance Inference Library for Transformers

Transformer, BERT and their variants have achieved great success in natural language processing. Since Transformer models are huge in size, serving these models is a challenge for real industrial applications. In this paper, we propose LightSeq, a highly efficient inference library for models in the Transformer family. LightSeq includes a series of GPU optimization techniques to to streamline the computation of neural layers and to reduce memory footprint. LightSeq can easily import models trained using PyTorch and Tensorflow. Experimental results on machine translation benchmarks show that LightSeq achieves up to 14x speedup compared with TensorFlow and 1.4x compared with FasterTransformer, a concurrent CUDA implementation. The code is available at https://github.com/bytedance/lightseq.

cs.MS↗

arXiv2021

How Information Diffuse in a Nomination Network

During the special period of the COVID-19 outbreak, this project investigated the driving factors in different information diffusion modes (i.e. broadcasting mode, contagion mode) based on the nomination relations in a social welfare campaign on Weibo. Specifically, we mapped a nomination social network and tracked the core communicators in both modes. Besides, we also observed the network from perspectives such as relationships between core communicators and modularity of the whole network. We extracted 6 homophily factors and tested them on 2 representative communities within the largest component of the network. We found that some core communicators distributed in a co-dependent way. At last, we supposed several explanations to the phenomenon which can be explored in further research.

cs.SI↗

arXiv2020

Ultra-uniform Nanocrystalline Materials via Two-Step Sintering

Nanocrystalline metals and ceramics with <100 nm grain sizes and superior properties (e.g., mechanical strength, hardness, fracture toughness and stored dielectric energy) are of great interest. Much has been discussed about achieving nano grains, but little is known about maintaining grain-size uniformity that is critical for material reliability. An especially intriguing question is whether it is possible to achieve a size distribution narrower than what Hillert[1] theoretically predicted for normal grain growth, a possibility suggested, for growth with a higher growth exponent, by the generalized mean-field theory[2] of Lifshitz, Slyozov, Wagner (LSW)[3,4] and Hillert but never realized in practice. We demonstrate that this can be achieved in bulk materials with an appropriately designed two-step sintering route that (a) takes advantage of the large growth exponent in the intermediate sintering stage to form a most uniform microstructure despite porosity remaining, and (b) freezes the grain growth thereon while continuing densification to reach full density. The resultant dense bulk Al2O3 ceramic has an average grain size of 34 nm and a much narrower size distribution than Hillert's prediction. Bulk Al2O3 with a grain-size distribution narrower than the particle-size distribution of starting powders was also demonstrated using this strategy, as were highly uniform bulk engineering metals and ceramics of either high purity and high melting points (Mo and W-Re) or highly complex compositions (core-shell BaTiO3 and 0.87BaTiO3-0.13Bi(Zn2/3(Nb0.85Ta0.15)1/3)O3).

cond-mat.mtrl-sci↗

arXiv2020

Multi-Task Reinforcement Learning based Mobile Manipulation Control for Dynamic Object Tracking and Grasping

Agile control of mobile manipulator is challenging because of the high complexity coupled by the robotic system and the unstructured working environment. Tracking and grasping a dynamic object with a random trajectory is even harder. In this paper, a multi-task reinforcement learning-based mobile manipulation control framework is proposed to achieve general dynamic object tracking and grasping. Several basic types of dynamic trajectories are chosen as the task training set. To improve the policy generalization in practice, random noise and dynamics randomization are introduced during the training process. Extensive experiments show that our policy trained can adapt to unseen random dynamic trajectories with about 0.1m tracking error and 75\% grasping success rate of dynamic objects. The trained policy can also be successfully deployed on a real mobile manipulator.

cs.RO↗

arXiv2020

An Efficient Agreement Mechanism in CapsNets By Pairwise Product

Capsule networks (CapsNets) are capable of modeling visual hierarchical relationships, which is achieved by the "routing-by-agreement" mechanism. This paper proposes a pairwise agreement mechanism to build capsules, inspired by the feature interactions of factorization machines (FMs). The proposed method has a much lower computation complexity. We further proposed a new CapsNet architecture that combines the strengths of residual networks in representing low-level visual features and CapsNets in modeling the relationships of parts to wholes. We conduct comprehensive experiments to compare the routing algorithms, including dynamic routing, EM routing, and our proposed FM agreement, based on both architectures of original CapsNet and our proposed one, and the results show that our method achieves both excellent performance and efficiency under a variety of situations.

cs.CV↗

arXiv2019

Directed momentum current induced by the PT-symmetric driving

We investigate the directed momentum current in the quantum kicked rotor model with $\mathcal{PT}$ symmetric deriving potential. For the quantum non-resonance case, the values of quasi-energy become to be complex when the strength of imaginary part of the kicking potential exceeds \textbf{a} threshold value, which demonstrates the appearance of the spontaneous $\mathcal{PT}$ symmetry breaking. In the vicinity of the phase-transition point, the momentum current exhibits a staircase growth with time. Each platform of the momentum current corresponds to the mean momentum of some eigenstates of the Floquet operator whose imaginary parts of the quasi-energy are significantly large. Above the phase-transition point, the momentum current increases linearly with time. Interestingly, its acceleration rate exhibits a kind of "quantized" increment with the kicking strength. We propose a modified classical acceleration mode of the kicked rotor model to explain such an intriguing phenomenon. Our theoretical prediction is in good agreement with numerical results.

quant-ph↗

arXiv2018

A theoretical interpretation of variance-based convergence criteria in perturbation-based theories

In QM/MM indirect free energy simulation, QM/MM corrections can be obtained from integration of partial derivatives of alchemical Hamiltonians or from perturbation-based estimators including free energy perturbation (FEP) and acceptance ratio methods. With FEP or exponential averaging, researchers tend to only sample MM states and calculate single point energy to get the free energy estimates. In this case the sample size hysteresis arises and the convergence is determined by bias elimination rather than variance minimization. Various criteria are proposed to evaluate the convergence issue and numerical studies are reported. It has been found that criteria including variance of distribution, effective sample size, information entropies and so on can be used and they are variance-of-distribution-dependent. However, no theoretical interpretation is presented. In this paper we present theoretical interpretations to dig the underlying statistical nature behind the problem. The convergence criteria are proven to be related with variance of distribution in Gaussian approximated Exponential averaging. Further, we prove that these estimators are nonlinearly dependent on the variance of the free energy estimate. As these estimators are often orders of magnitude overestimated, the variance of the FEP estimate is orders of magnitude underestimated. Hence, computing this statistical uncertainty is meaningless. In numerical calculation from timeseries data the effective sample size is bounded by 1 and N and thus the variance of the free energy estimate is proven to be bounded by 0 and 1 (kBT)2 for EXP and 0 and 2 (kBT)2 for BAR, which indicates an inevitable underestimation. Specifically, the upper bounds for these estimators are sample-size dependent. The effective sample size is proven to be a function of the overlap scalar, from which the range of the overlap scalar can also be derived.

physics.chem-ph↗

arXiv2018

Image Domain Dual Material Decomposition for Dual-Energy CT using Butterfly Network

Dual-energy CT (DECT) has been increasingly used in imaging applications because of its capability for material differentiation. However, material decomposition suffers from magnified noise from two CT images of independent scans, leading to severe degradation of image quality. Existing algorithms achieve suboptimal decomposition performance since they fail to accurately depict the mapping relationship between DECT and the basis material images. Inspired by the impressive potential of CNN, we developed a new Butterfly network to perform the image domain dual material decomposition due to its strong approximation ability to the mapping functions in DECT. The Butterfly network is derived from the image domain DECT decomposition model by exploring the geometric relationship between mapping functions of data model and network components. The network is designed as the double-entry double-out crossover architecture based on the decomposition formulation. It enters a pair of dual-energy images as inputs and defines the ground true decomposed images as each label. The crossover architecture, which plays an important role in material decomposition, is designed to implement the information exchange between the two material generation pathways in the network. Network components exhibit different sensitivity to basis materials in the visualization. By analyzing their sensitivity to different basis materials, we determined the roles of network components in material decomposition. This visualization evaluation reveals what the network can learn and verifies the rationality of network design. The qualitative and quantitative evaluations in material decomposition of patient data indicate that the proposed network outperforms its counterpart.

physics.med-ph↗

arXiv2018

An efficient global optimization algorithm for maximizing the sum of two generalized Rayleigh quotients

Maximizing the sum of two generalized Rayleigh quotients (SRQ) can be reformulated as a one-dimensional optimization problem, where the function value evaluations are reduced to solving semi-definite programming (SDP) subproblems. In this paper, we first use the dual SDP subproblem to construct an explicit overestimation and then propose a branch-and-bound algorithm to globally solve (SRQ). Numerical results demonstrate that it is even more efficient than the recent SDP-based heuristic algorithm.

math.OC↗

arXiv2017

The Plastic Scintillator Detector at DAMPE

he DArk Matter Particle Explorer (DAMPE) is a general purposed satellite-borne high energy $γ-$ray and cosmic ray detector, and among the scientific objectives of DAMPE are the searches for the origin of cosmic rays and an understanding of Dark Matter particles. As one of the four detectors in DAMPE, the Plastic Scintillator Detector (PSD) plays an important role in the particle charge measurement and the photons/electrons separation. The PSD has 82 modules, each consists of a long organic plastic scintillator bar and two PMTs at both ends for readout, in two layers and covers an overall active area larger than 82 cm $\times$ 82 cm. It can identify the charge states for relativistic ions from H to Fe, and the detector efficiency for Z=1 particles can reach 0.9999. The PSD has been successfully launched with DAMPE on Dec. 17, 2015. In this paper, the design, the assembly, the qualification tests of the PSD and some of the performance measured on the ground have been described in detail.

astro-ph.IM↗

arXiv2016

Comment on "Anomalous Edge State in a Non-Hermitian Lattice"

In this comment, we criticize three main conclusions of the letter\cite{Lee2016}. We show that the concept of fractional winding number(FWN) is factitious, Lee's conclusions on Fig. 3 are finite-size effect and the breakdown of bulk-boundary correspondence (BBBC) cannot be explained by "defective".

quant-ph↗