arXiv subjects

Peng Chen

Publications and source records attributed to Peng Chen.

At least 325 records · Page 18Linked to original sources

arXiv2021

FATNN: Fast and Accurate Ternary Neural Networks

Ternary Neural Networks (TNNs) have received much attention due to being potentially orders of magnitude faster in inference, as well as more power efficient, than full-precision counterparts. However, 2 bits are required to encode the ternary representation with only 3 quantization levels leveraged. As a result, conventional TNNs have similar memory consumption and speed compared with the standard 2-bit models, but have worse representational capability. Moreover, there is still a significant gap in accuracy between TNNs and full-precision networks, hampering their deployment to real applications. To tackle these two challenges, in this work, we first show that, under some mild constraints, computational complexity of the ternary inner product can be reduced by a factor of 2. Second, to mitigate the performance gap, we elaborately design an implementation-dependent ternary quantization algorithm. The proposed framework is termed Fast and Accurate Ternary Neural Networks (FATNN). Experiments on image classification demonstrate that our FATNN surpasses the state-of-the-arts by a significant margin in accuracy. More importantly, speedup evaluation compared with various precisions is analyzed on several platforms, which serves as a strong benchmark for further research.

cs.LG↗

arXiv2021

ABCNet v2: Adaptive Bezier-Curve Network for Real-time End-to-end Text Spotting

End-to-end text-spotting, which aims to integrate detection and recognition in a unified framework, has attracted increasing attention due to its simplicity of the two complimentary tasks. It remains an open problem especially when processing arbitrarily-shaped text instances. Previous methods can be roughly categorized into two groups: character-based and segmentation-based, which often require character-level annotations and/or complex post-processing due to the unstructured output. Here, we tackle end-to-end text spotting by presenting Adaptive Bezier Curve Network v2 (ABCNet v2). Our main contributions are four-fold: 1) For the first time, we adaptively fit arbitrarily-shaped text by a parameterized Bezier curve, which, compared with segmentation-based methods, can not only provide structured output but also controllable representation. 2) We design a novel BezierAlign layer for extracting accurate convolution features of a text instance of arbitrary shapes, significantly improving the precision of recognition over previous methods. 3) Different from previous methods, which often suffer from complex post-processing and sensitive hyper-parameters, our ABCNet v2 maintains a simple pipeline with the only post-processing non-maximum suppression (NMS). 4) As the performance of text recognition closely depends on feature alignment, ABCNet v2 further adopts a simple yet effective coordinate convolution to encode the position of the convolutional filters, which leads to a considerable improvement with negligible computation overhead. Comprehensive experiments conducted on various bilingual (English and Chinese) benchmark datasets demonstrate that ABCNet v2 can achieve state-of-the-art performance while maintaining very high efficiency.

cs.CV↗

arXiv2021

Sharp endpoint estimates for Schrödinger groups on Hardy spaces

Let $L$ be a non-negative self-adjoint operator acting on $L^2(X)$ where $X$ is a space of homogeneous type with a dimension $n$. Suppose that the heat kernel of $L$ satisfies the Davies-Gaffney estimates of order $m\geq 2$. Let $H^1_L(X)$ be the Hardy space associated with $L.$ In this paper we show sharp endpoint estimate for the Schrödinger group $e^{itL}$ associated with $L$ such that \begin{eqnarray*} \left\| (I+L)^{-{n/2}}e^{itL} f\right\|_{ L^1(X)} + \left\| (I+L)^{-{n/2}}e^{itL} f\right\|_{ H^1_L(X)} \leq C(1+|t|)^{n/2}\|f\|_{H^1_L(X)}, \ \ \ t\in{\mathbb R} \end{eqnarray*} for some constant $C=C(n, m)>0$ independent of $t$. By a duality and interpolation argument, it gives a new proof of a recent result of \cite{CDLY} for { sharp} endpoint $L^p$-Sobolev bound for $e^{itL}$: $$ \left\| (I+L)^{-s }e^{itL} f\right\|_{ L^p(X)} \leq C (1+|t|)^{s} \|f\|_{ L^p(X)}, \ \ \ t\in{\mathbb R}, \ \ \ s\geq n\big|{1\over 2}-{1\over p}\big| $$ for every $1<p<\infty$ when the heat kernel of $L$ satisfies a Gaussian upper bound, which extends the classical results due to Miyachi ) for the Laplacian on the Euclidean space ${\mathbb R}^n$.

math.AP↗

arXiv2021

Smoothed Multi-View Subspace Clustering

In recent years, multi-view subspace clustering has achieved impressive performance due to the exploitation of complementary imformation across multiple views. However, multi-view data can be very complicated and are not easy to cluster in real-world applications. Most existing methods operate on raw data and may not obtain the optimal solution. In this work, we propose a novel multi-view clustering method named smoothed multi-view subspace clustering (SMVSC) by employing a novel technique, i.e., graph filtering, to obtain a smooth representation for each view, in which similar data points have similar feature values. Specifically, it retains the graph geometric features through applying a low-pass filter. Consequently, it produces a ``clustering-friendly" representation and greatly facilitates the downstream clustering task. Extensive experiments on benchmark datasets validate the superiority of our approach. Analysis shows that graph filtering increases the separability of classes.

cs.CV↗

arXiv2021

Active control of a diode laser with injection locking

We present a simple and effective method to implement an active stabilization of a diode laser with injection locking, which requires minimal user intervenes. The injection locked state of the diode laser is probed by a photodetector, of which sensitivity is enhanced by a narrow laser-line filter. Taking advantage of the characteristic response of laser power to spectral modes from the narrow laser-line filter, we demonstrate that high spectral purity and low intensity noise of the diode can be simultaneously maintained by an active feedback to the injected laser. Our method is intrinsically cost-effective, and does not require bulky devices, such as Fabry-Perot interferometers or wavemeters, to actively stabilize the diode laser. Based on successful implementation of this method in our quantum gas experiments, it is conceivable that our active stabilization will greatly simplify potential applications of injection locking of diode lasers in modularized or integrated optical systems.

physics.optics↗

arXiv2021

Thickness-Driven Quantum Anomalous Hall Phase Transition in Magnetic Topological Insulator Thin Films

The quantized version of anomalous Hall effect realized in magnetic topological insulators (MTIs) has great potential for the development of topological quantum physics and low-power electronic/spintronic applications. To enable dissipationless chiral edge conduction at zero magnetic field, effective exchange field arisen from the aligned magnetic dopants needs to be large enough to yield specific spin sub-band configurations. Here we report the thickness-tailored quantum anomalous Hall (QAH) effect in Cr-doped (Bi,Sb)2Te3 thin films by tuning the system across the two-dimensional (2D) limit. In addition to the Chern number-related metal-to-insulator QAH phase transition, we also demonstrate that the induced hybridization gap plays an indispensable role in determining the ground magnetic state of the MTIs, namely the spontaneous magnetization owning to considerable Van Vleck spin susceptibility guarantees the zero-field QAH state with unitary scaling law in thick samples, while the quantization of the Hall conductance can only be achieved with the assistance of external magnetic fields in ultra-thin films. The modulation of topology and magnetism through structural engineering may provide a useful guidance for the pursuit of QAH-based new phase diagrams and functionalities.

cond-mat.mes-hall↗

arXiv2021

Performance Portable Back-projection Algorithms on CPUs: Agnostic Data Locality and Vectorization Optimizations

Computed Tomography (CT) is a key 3D imaging technology that fundamentally relies on the compute-intense back-projection operation to generate 3D volumes. GPUs are typically used for back-projection in production CT devices. However, with the rise of power-constrained micro-CT devices, and also the emergence of CPUs comparable in performance to GPUs, back-projection for CPUs could become favorable. Unlike GPUs, extracting parallelism for back-projection algorithms on CPUs is complex given that parallelism and locality are not explicitly defined and controlled by the programmer, as is the case when using CUDA for instance. We propose a collection of novel back-projection algorithms that reduce the arithmetic computation, robustly enable vectorization, enforce a regular memory access pattern, and maximize the data locality. We also implement the novel algorithms as efficient back-projection kernels that are performance portable over a wide range of CPUs. Performance evaluation using a variety of CPUs from different vendors and generations demonstrates that our back-projection implementation achieves on average 5.2x speedup over the multi-threaded implementation of the most widely used, and optimized, open library. With a state-of-the-art CPU, we reach performance that rivals top-performing GPUs.

cs.DC↗

arXiv2021

Physics-Based Iterative Reconstruction for Dual Source and Flying Focal Spot Computed Tomography

For single source helical Computed Tomography (CT), both Filtered-Back Projection (FBP) and statistical iterative reconstruction have been investigated. However for dual source CT with flying focal spot (DS-FFS CT), statistical iterative reconstruction that accurately models the scanner geometry and physics remains unknown to researchers. Therefore, this paper presents a novel physics-based iterative reconstruction method for DS-FFS CT and assess its image quality. Our algorithm uses precise physics models to reconstruct from the native cone-beam geometry and interleaved dual source helical trajectory of a DS-FFS CT. To do so, we construct a noise physics model to represent data acquisition noise and a prior image model to represent image noise and texture. In addition, we design forward system models to compute the locations of deflected focal spots, the dimension and sensitivity of voxels and detector units, as well as the length of intersection between X-rays and voxels. The forward system models further represent the coordinated movement between the dual sources by computing their X-ray coverage gaps and overlaps at an arbitrary helical pitch. With the above models, we reconstruct images by using an advanced Consensus Equilibrium (CE) numerical method to compute the maximum a posteriori estimate to a joint optimization problem that simultaneously fits all models. We compared our reconstruction with Siemens ADMIRE, which is the clinical standard hybrid iterative reconstruction (IR) method for DS-FFS CT, in terms of spatial resolution, noise profile and image artifacts through both phantoms and clinical datasets. Experiments show that our reconstruction has a consistently higher spatial resolution than the clinical standard hybrid IR. In addition, our reconstruction shows a reduced magnitude of image undersampling artifacts than the clinical standard.

eess.IV↗

arXiv2021

Fully Quantized Image Super-Resolution Networks

With the rising popularity of intelligent mobile devices, it is of great practical significance to develop accurate, realtime and energy-efficient image Super-Resolution (SR) inference methods. A prevailing method for improving the inference efficiency is model quantization, which allows for replacing the expensive floating-point operations with efficient fixed-point or bitwise arithmetic. To date, it is still challenging for quantized SR frameworks to deliver feasible accuracy-efficiency trade-off. Here, we propose a Fully Quantized image Super-Resolution framework (FQSR) to jointly optimize efficiency and accuracy. In particular, we target on obtaining end-to-end quantized models for all layers, especially including skip connections, which was rarely addressed in the literature. We further identify training obstacles faced by low-bit SR networks and propose two novel methods accordingly. The two difficulites are caused by 1) activation and weight distributions being vastly distinctive in different layers; 2) the inaccurate approximation of the quantization. We apply our quantization scheme on multiple mainstream super-resolution architectures, including SRResNet, SRGAN and EDSR. Experimental results show that our FQSR using low bits quantization can achieve on par performance compared with the full-precision counterparts on five benchmark datasets and surpass state-of-the-art quantized SR methods with significantly reduced computational cost and memory consumption.

eess.IV↗

arXiv2021

LI-Net: Large-Pose Identity-Preserving Face Reenactment Network

Face reenactment is a challenging task, as it is difficult to maintain accurate expression, pose and identity simultaneously. Most existing methods directly apply driving facial landmarks to reenact source faces and ignore the intrinsic gap between two identities, resulting in the identity mismatch issue. Besides, they neglect the entanglement of expression and pose features when encoding driving faces, leading to inaccurate expressions and visual artifacts on large-pose reenacted faces. To address these problems, we propose a Large-pose Identity-preserving face reenactment network, LI-Net. Specifically, the Landmark Transformer is adopted to adjust driving landmark images, which aims to narrow the identity gap between driving and source landmark images. Then the Face Rotation Module and the Expression Enhancing Generator decouple the transformed landmark image into pose and expression features, and reenact those attributes separately to generate identity-preserving faces with accurate expressions and poses. Both qualitative and quantitative experimental results demonstrate the superiority of our method.

cs.CV↗

arXiv2021

Robust optimization Design of a New Combined Median Barrier Based on Taguchi method and Grey Relational Analysis

Accidents that vehicles cross median and enter opposite lane happen frequently, and the existing median barrier has weak anti-collision strength. A new combined median barrier (NCMB) consisted of W-beam guardrail and concrete structure was proposed to decrease deformation and enhance anti-collision strength in this paper. However, there were some uncertainties in the initial design of the NCMB. If the uncertainties were not considered in the design process, the optimization objectives were especially sensitive to the small fluctuation of the variables, and it might result in design failure. For this purpose, the acceleration and deflection were taken as objectives; post thickness, W-beam thickness and post spacing were chosen as design variables; the velocity, mass of vehicle and the yield stress of barrier components were taken as noise factors, a multi-objective robust optimization is carried out for the NCMB based on Taguchi and grey relational analysis (GRA). The results indicate that the acceleration and deflection after optimization are reduced by 47.3% and 76.7% respectively; Signal-to-noise ratio (SNR) of objectives after optimization are increased, it greatly enhances the robustness of the NCMB. The results demonstrate that the effectiveness of the methodology that based on Taguchi method and grey relational analysis.

math.NA↗

arXiv2021

xERTE: Explainable Reasoning on Temporal Knowledge Graphs for Forecasting Future Links

Modeling time-evolving knowledge graphs (KGs) has recently gained increasing interest. Here, graph representation learning has become the dominant paradigm for link prediction on temporal KGs. However, the embedding-based approaches largely operate in a black-box fashion, lacking the ability to interpret their predictions. This paper provides a link forecasting framework that reasons over query-relevant subgraphs of temporal KGs and jointly models the structural dependencies and the temporal dynamics. Especially, we propose a temporal relational attention mechanism and a novel reverse representation update scheme to guide the extraction of an enclosing subgraph around the query. The subgraph is expanded by an iterative sampling of temporal neighbors and by attention propagation. Our approach provides human-understandable evidence explaining the forecast. We evaluate our model on four benchmark temporal knowledge graphs for the link forecasting task. While being more explainable, our model obtains a relative improvement of up to 20% on Hits@1 compared to the previous best KG forecasting method. We also conduct a survey with 53 respondents, and the results show that the evidence extracted by the model for link forecasting is aligned with human understanding.

cs.LG↗

arXiv2021

Derivative-Informed Projected Neural Networks for High-Dimensional Parametric Maps Governed by PDEs

Many-query problems, arising from uncertainty quantification, Bayesian inversion, Bayesian optimal experimental design, and optimization under uncertainty-require numerous evaluations of a parameter-to-output map. These evaluations become prohibitive if this parametric map is high-dimensional and involves expensive solution of partial differential equations (PDEs). To tackle this challenge, we propose to construct surrogates for high-dimensional PDE-governed parametric maps in the form of projected neural networks that parsimoniously capture the geometry and intrinsic low-dimensionality of these maps. Specifically, we compute Jacobians of these PDE-based maps, and project the high-dimensional parameters onto a low-dimensional derivative-informed active subspace; we also project the possibly high-dimensional outputs onto their principal subspace. This exploits the fact that many high-dimensional PDE-governed parametric maps can be well-approximated in low-dimensional parameter and output subspace. We use the projection basis vectors in the active subspace as well as the principal output subspace to construct the weights for the first and last layers of the neural network, respectively. This frees us to train the weights in only the low-dimensional layers of the neural network. The architecture of the resulting neural network captures to first order, the low-dimensional structure and geometry of the parametric map. We demonstrate that the proposed projected neural network achieves greater generalization accuracy than a full neural network, especially in the limited training data regime afforded by expensive PDE-based parametric maps. Moreover, we show that the number of degrees of freedom of the inner layers of the projected network is independent of the parameter and output dimensions, and high accuracy can be achieved with weight dimension independent of the discretization dimension.

math.NA↗

arXiv2021

Towards Enhancing Database Education: Natural Language Generation Meets Query Execution Plans

The database systems course is offered as part of an undergraduate computer science degree program in many major universities. A key learning goal of learners taking such a course is to understand how SQL queries are processed in a RDBMS in practice. Since a query execution plan (QEP) describes the execution steps of a query, learners can acquire the understanding by perusing the QEPs generated by a RDBMS. Unfortunately, in practice, it is often daunting for a learner to comprehend these QEPs containing vendor-specific implementation details, hindering her learning process. In this paper, we present a novel, end-to-end, generic system called lantern that generates a natural language description of a qep to facilitate understanding of the query execution steps. It takes as input an SQL query and its QEP, and generates a natural language description of the execution strategy deployed by the underlying RDBMS. Specifically, it deploys a declarative framework called pool that enables subject matter experts to efficiently create and maintain natural language descriptions of physical operators used in QEPs. A rule-based framework called RULE-LANTERN is proposed that exploits pool to generate natural language descriptions of QEPs. Despite the high accuracy of RULE-LANTERN, our engagement with learners reveal that, consistent with existing psychology theories, perusing such rule-based descriptions lead to boredom due to repetitive statements across different QEPs. To address this issue, we present a novel deep learning-based language generation framework called NEURAL-LANTERN that infuses language variability in the generated description by exploiting a set of paraphrasing tools and word embedding. Our experimental study with real learners shows the effectiveness of lantern in facilitating comprehension of QEPs.

cs.DB↗

arXiv2021

Matrix Engines for High Performance Computing:A Paragon of Performance or Grasping at Straws?

Matrix engines or units, in different forms and affinities, are becoming a reality in modern processors; CPUs and otherwise. The current and dominant algorithmic approach to Deep Learning merits the commercial investments in these units, and deduced from the No.1 benchmark in supercomputing, namely High Performance Linpack, one would expect an awakened enthusiasm by the HPC community, too. Hence, our goal is to identify the practical added benefits for HPC and machine learning applications by having access to matrix engines. For this purpose, we perform an in-depth survey of software stacks, proxy applications and benchmarks, and historical batch job records. We provide a cost-benefit analysis of matrix engines, both asymptotically and in conjunction with state-of-the-art processors. While our empirical data will temper the enthusiasm, we also outline opportunities to misuse these dense matrix-multiplication engines if they come for free.

cs.DC↗

arXiv2021

Projected Wasserstein gradient descent for high-dimensional Bayesian inference

We propose a projected Wasserstein gradient descent method (pWGD) for high-dimensional Bayesian inference problems. The underlying density function of a particle system of WGD is approximated by kernel density estimation (KDE), which faces the long-standing curse of dimensionality. We overcome this challenge by exploiting the intrinsic low-rank structure in the difference between the posterior and prior distributions. The parameters are projected into a low-dimensional subspace to alleviate the approximation error of KDE in high dimensions. We formulate a projected Wasserstein gradient flow and analyze its convergence property under mild assumptions. Several numerical experiments illustrate the accuracy, convergence, and complexity scalability of pWGD with respect to parameter dimension, sample size, and processor cores.

cs.LG↗

arXiv2021

$SU(5)$ GUTs with $A_4$ modular symmetry

We combine $SU(5)$ Grand Unified Theories (GUTs) with $A_4$ modular symmetry and present a comprehensive analysis of the resulting quark and lepton mass matrices for all the simplest cases. Classifying the models according to the representation assignments of the matter fields under $A_4$, we find that there are seven types of $SU(5)$ models with $A_4$ modular symmetry. We present 53 benchmark models with the fewest free parameters. The parameter space of each model is scanned to optimize the agreement between predictions and experimental data, and predictions for the masses and mixing parameters of quarks and leptons are given at the best fitting points. The best fit predictions for the leptonic CP violating Dirac phase, the lightest neutrino mass and the neutrinoless double beta decay parameter when displayed graphically are observed to cover a wide range of possible values, but are clustered around particular regions, allowing future neutrino experiments to discriminate between the different types of models.

hep-ph↗

arXiv2021

Quantum teleportation mediated by surface plasmon polariton

Surface plasmon polaritons (SPPs) are collective excitations of free electrons propagating along a metal-dielectric interface. Although some basic quantum properties of SPPs, such as the preservation of entanglement, the wave-particle duality of a single plasmon, the quantum interference of two plasmons, and the verification of entanglement generation, have been shown, more advanced quantum information protocols have yet to be demonstrated with SPPs. Here, we experimentally realize quantum state teleportation between single photons and SPPs. To achieve this, we use polarization-entangled photon pairs, coherent photon-plasmon-photon conversion on a metallic subwavelength hole array, complete Bell-state measurements and an active feed-forward technique. The results of both quantum state and quantum process tomography confirm the quantum nature of the SPP mediated teleportation. An average state fidelity of 0.889$\pm$0.004 and a process fidelity of 0.820$\pm$0.005, which are well above the classical limit, are achieved. Our work shows that SPPs may be useful for realizing complex quantum protocols in a photonic-plasmonic hybrid quantum network.

quant-ph↗