arXiv subjects

Yu Zhu

Publications and source records attributed to Yu Zhu.

At least 163 records · Page 9Linked to original sources

arXiv2018

A Novel Co-design Peta-scale Heterogeneous Cluster for Deep Learning Training

Large scale deep Convolution Neural Networks (CNNs) increasingly demands the computing power. It is key for researchers to own a great powerful computing platform to leverage deep learning (DL) advancing.On the other hand, as the commonly-used accelerator, the commodity GPUs cards of new generations are more and more expensive. Consequently, it is of importance to design an affordable distributed heterogeneous system that provides powerful computational capacity and develop a well-suited software that efficiently utilizes its computational capacity. In this paper, we present our co-design distributed system including a peta-scale GPU cluster, called "Manoa". Based on properties and topology of Manoa, we first propose job server framework and implement it, named "MiMatrix". The central node of MiMatrix, referred to as the job server, undertakes all of controlling, scheduling and monitoring, and I/O tasks without weight data transfer for AllReduce processing in each iteration. Therefore, MiMatrix intrinsically solves the bandwidth bottleneck of central node in parameter server framework that is widely used in distributed DL tasks. Meanwhile, we also propose a new AllReduce algorithm, GPUDirect RDMA-Aware AllReduce~(GDRAA), in which both computation and handshake message are O(1) and the number of synchronization is two in each iteration that is a theoretical minimum number. Owe to the dedicated co-design distributed system, MiMatrix efficiently makes use of the Manoa's computational capacity and bandwidth. We benchmark Manoa Resnet50 and Resenet101 on Imagenet-1K dataset. Some of results have demonstrated state-of-the-art.

cs.CV↗

arXiv2017

ChineseFoodNet: A large-scale Image Dataset for Chinese Food Recognition

In this paper, we introduce a new and challenging large-scale food image dataset called "ChineseFoodNet", which aims to automatically recognizing pictured Chinese dishes. Most of the existing food image datasets collected food images either from recipe pictures or selfie. In our dataset, images of each food category of our dataset consists of not only web recipe and menu pictures but photos taken from real dishes, recipe and menu as well. ChineseFoodNet contains over 180,000 food photos of 208 categories, with each category covering a large variations in presentations of same Chinese food. We present our efforts to build this large-scale image dataset, including food category selection, data collection, and data clean and label, in particular how to use machine learning methods to reduce manual labeling work that is an expensive process. We share a detailed benchmark of several state-of-the-art deep convolutional neural networks (CNNs) on ChineseFoodNet. We further propose a novel two-step data fusion approach referred as "TastyNet", which combines prediction results from different CNNs with voting method. Our proposed approach achieves top-1 accuracies of 81.43% on the validation set and 81.55% on the test set, respectively. The latest dataset is public available for research and can be achieved at https://sites.google.com/view/chinesefoodnet.

cs.CV↗

arXiv2016

Strong interfacial exchange field in the graphene/EuS heterostructure

Exploiting 2D materials for spintronic applications can potentially realize next-generation devices featuring low-power consumption and quantum operation capability. The magnetic exchange field (MEF) induced by an adjacent magnetic insulator enables efficient control of local spin generation and spin modulation in 2D devices without compromising the delicate material structures. Using graphene as a prototypical 2D system, we demonstrate that its coupling to the model magnetic insulator (EuS) produces a substantial MEF (> 14 T) with potential to reach hundreds of Tesla, which leads to orders-of-magnitude enhancement in the spin signal originated from Zeeman spin-Hall effect. Furthermore, the new ferromagnetic ground state of Dirac electrons resulting from the strong MEF may give rise to quantized spin-polarized edge transport. The MEF effect shown in our graphene/EuS devices therefore provides a key functionality for future spin logic and memory devices based on emerging 2D materials in classical and quantum information processing.

cond-mat.mes-hall↗

arXiv2015

Fidelity-Naturalness Evaluation of Single Image Super Resolution

We study the problem of evaluating super resolution methods. Traditional evaluation methods usually judge the quality of super resolved images based on a single measure of their difference with the original high resolution images. In this paper, we proposed to use both fidelity (the difference with original images) and naturalness (human visual perception of super resolved images) for evaluation. For fidelity evaluation, a new metric is proposed to solve the bias problem of traditional evaluation. For naturalness evaluation, we let humans label preference of super resolution results using pair-wise comparison, and test the correlation between human labeling results and image quality assessment metrics' outputs. Experimental results show that our fidelity-naturalness method is better than the traditional evaluation method for super resolution methods, which could help future research on single-image super resolution.

cs.CV↗

arXiv2015

Model-Free Sure Screening via Maximum Correlation

We consider the problem of screening features in an ultrahigh-dimensional setting. Using maximum correlation, we develop a novel procedure called MC-SIS for feature screening, and show that MC-SIS possesses the sure screen property without imposing model or distributional assumptions on the response and predictor variables. Therefore, MC-SIS is a model-free sure independence screening method as in contrast with some other existing model-based sure independence screening methods in the literature. Simulation examples and a real data application are used to demonstrate the performance of MC-SIS as well as to compare MC-SIS with other existing sure screening methods. The results show that MC-SIS outperforms those methods when their model assumptions are violated, and it remains competitive when the model assumptions hold.

stat.ME↗

arXiv2015

Spin-Polarized Tunneling Study on Spin-Momentum Locking in the Topological Insulators

We demonstrate that the charge-spin conversion efficiency of topological insulators (TI) can be experimentally determined by injecting spin-polarized tunneling electrons into a TI. Through a comparative study between bismuth selenide and bismuth antimony telluride, we verified the topological-surface-state origin of the observed giant spin signals. By injecting energetic electrons into bismuth selenide, we further studied the energy dependence of the effective spin polarization at the TI surface. The experimentally verified large spin polarization, as well as our calculations, provides new insights into optimizing TI materials for near room-temperature spintronic applications.

cond-mat.mes-hall↗

arXiv2014

Statistical calibration of qRT-PCR, microarray and RNA-Seq gene expression data with measurement error models

The accurate quantification of gene expression levels is crucial for transcriptome study. Microarray platforms are commonly used for simultaneously interrogating thousands of genes in the past decade, and recently RNA-Seq has emerged as a promising alternative. The gene expression measurements obtained by microarray and RNA-Seq are, however, subject to various measurement errors. A third platform called qRT-PCR is acknowledged to provide more accurate quantification of gene expression levels than microarray and RNA-Seq, but it has limited throughput capacity. In this article, we propose to use a system of functional measurement error models to model gene expression measurements and calibrate the microarray and RNA-Seq platforms with qRT-PCR. Based on the system, a two-step approach was developed to estimate the biases and error variance components of the three platforms and calculate calibrated estimates of gene expression levels. The estimated biases and variance components shed light on the relative strengths and weaknesses of the three platforms and the calibrated estimates provide a more accurate and consistent quantification of gene expression levels. Theoretical and simulation studies were conducted to establish the properties of those estimates. The system was applied to analyze two gene expression data sets from the Microarray Quality Control (MAQC) and Sequencing Quality Control (SEQC) projects.

stat.AP↗

arXiv2014

Direct tunneling through high-$κ$ amorphous HfO$_2$: effects of chemical modification

We report first principles modeling of quantum tunneling through amorphous HfO$_2$ dielectric layer of metal-oxide-semiconductor (MOS) nanostructures in the form of n-Si/HfO$_2$/Al. In particular we predict that chemically modifying the amorphous HfO$_2$ barrier by doping N and Al atoms in the middle region - far from the two interfaces of the MOS structure, can reduce the gate-to-channel tunnel leakage by more than one order of magnitude. Several other types of modification are found to enhance tunneling or induce substantial band bending in the Si, both are not desired from leakage point of view. By analyzing transmission coefficients and projected density of states, the microscopic physics of electron traversing the tunnel barrier with or without impurity atoms in the high-$κ$ dielectric is revealed.

cond-mat.mtrl-sci↗

arXiv2013

Nonequilibrium Green's function theory for predicting device-to-device variability

Due to random dopant fluctuations, the device-to-device variability is a serious challenge to emerging nanoelectronics. In this work we present theoretical formalisms and numerical simulations of quantum transport variability, based on the nonequilibrium Green's functions and the multiple scattering theory. We have developed a general formalism using the diagrammatic technique within the coherent potential approximation (CPA) that can be applied to a wide range of disorder concentrations. In addition, we have developed a method by using a perturbative expansion within the low concentration approximation (LCA) that is extremely useful for typical nanoelectronic devices having low dopant concentration. Applying both formalisms, transport fluctuations due to random impurities can be predicted without lengthy brute force computation of ensemble of devices structures. Numerical implementations of the formalisms are demonstrated using both tight-binding models and first principles models.

cond-mat.mes-hall↗

arXiv2013

Quantum Transport Theory with the Nonequilibrium Coherent Potentials

Since any realistic electronic device has some degree of disorder, predicting disorder effects in quantum transport is a critical problem. Here we report the theory of nonequilibrium coherent potential approximation (NECPA) for analyzing disorder effects in nonequilibrium quantum transport of nanoelectronic devices. The NECPA is formulated by contour ordered nonequilibrium Green's function (NEGF) where the disorder average is carried out within the coherent potential approximation on the complex-time contour. We have derived a set of new rules that supplement the celebrated Langreth theorem and, as a whole, the generalized Langreth rules allow us to derive NECPA equations for real time Green's functions. The solution of NECPA equations provide the disorder averaged nonequilibrium density matrix as well as other relevant quantities for quantum transport calculations. We establish the excellent accuracy of NECPA by comparing its results to brute force numerical calculations of disordered tight-binding models. Moreover, the connection of NECPA equations which are derived on the complex-time contour, to the nonequilibrium vertex correction theory which is derived on the real-time axis, is made. As an application, we demonstrate that NECPA can be combined with density functional theory to enable analysis of nanoelectronic device physics from atomistic first principles.

cond-mat.mes-hall↗

arXiv2013

Electric control of spin in monolayer WSe2 field effect transistors

We report a first principles theoretical investigation of quantum transport in monolayer WSe2 field effect transistor (FET). Due to a strong spin-orbit interaction (SOI) and the atomic structure of the two-dimensional (2D) lattice, monolayer WSe2 has an interesting electronic structure that exhibits Zeeman-like up-down spin texture near the K and K' points of the Brillouin zone. In a FET, the gate electric field induces an extra, externally tunable SOI that re-orients the spins into a Rashba-like texture thereby realizing electric control of the spin. Quantum transport is modulated by the spin texture, namely by if the spin orientation of the carrier after the gated channel region, matches or miss-matches that of the FET drain electrode. The carrier current in the FET is labelled both the spin index and the valley index, realizing spintronics and valleytronics in the same device.

cond-mat.mes-hall↗

arXiv2013

Band offset of GaAs/AlxGa1-xAs heterojunctions from atomistic first principles

Using an atomistic first principles approach, we investigate the band offset of the GaAs/AlxGa1-xAs heterojunctions for the entire range of the Al doping concentration 0<x<=1. We apply the coherent potential approach to handle the configuration average of Al doping and a recently proposed semi-local exchange potential to accurately determine the band gaps of the materials. The calculated band structures of the GaAs, AlAs crystals and band gaps of the GaAs/AlxGa1-xAs alloys, are in very good agreement with the experimental results. We predict that valence band offset of the GaAs/AlxGa1-xAs heterojunction scales with the Al concentration x in a linear fashion as VBO(x)~0.587 x, and the conduction band offset scales with x in a nonlinear fashion. Quantitative comparisons to the corresponding experimental data are made.

cond-mat.mtrl-sci↗

arXiv2013

A generic tight-binding model for monolayer, bilayer and bulk MoS2

Molybdenum disulfide (MoS2) is a layered semiconductor which has become very important recently as an emerging electronic device material. Being an intrinsic semiconductor the two-dimensional MoS2 has major advantages as the channel material in field-effect transistors. In this work we determine the electronic structure of MoS2 with the highly accurate screened hybrid functional within the density functional theory (DFT) including the spin-orbit coupling. Using the DFT electronic structures as target, we have developed a single generic tight-binding (TB) model that accurately produces the electronic structures for three different forms of MoS2 - bulk, bilayer and monolayer. Our TB model is based on the Slater-Koster method with non-orthogonal sp3d5 orbitals, nearest-neighbor interactions and spin-orbit coupling. The TB model is useful for atomistic modeling of quantum transport in MoS2 based electronic devices.

cond-mat.mes-hall↗

arXiv2013

Electronic structures of III-V zinc-blende semiconductors from atomistic first principles

For analyzing quantum transport in semiconductor devices, accurate electronic structures are critical for quantitative predictions. Here we report theoretical analysis of electronic structures of all III-V zinc-blende semiconductor compounds. Our calculations are from density functional theory with the semi-local exchange proposed recently [F. Tran and P. Blaha, Phys. Rev. Lett. 102, 226401 (2009)], within the linear muffin tin orbital scheme. The calculated band gaps and effective masses are compared to experimental data and good quantitative agreement is obtained. Using the theoretical scheme presented here, quantum transport in nanostructures of III-V compounds can be confidently predicted.

cond-mat.mtrl-sci↗

arXiv2012

Structure and electronic transport in graphene wrinkles

Wrinkling is a ubiquitous phenomenon in two-dimensional membranes. In particular, in the large-scale growth of graphene on metallic substrates, high densities of wrinkles are commonly observed. Despite their prevalence and potential impact on large-scale graphene electronics, relatively little is known about their structural morphology and electronic properties. Surveying the graphene landscape using atomic force microscopy, we found that wrinkles reach a certain maximum height before folding over. Calculations of the energetics explain the morphological transition, and indicate that the tall ripples are collapsed into narrow standing wrinkles by van der Waals forces, analogous to large-diameter nanotubes. Quantum transport calculations show that conductance through these collapsed wrinkle structures is limited mainly by a density-of-states bottleneck and by interlayer tunneling across the collapsed bilayer region. Also through systematic measurements across large numbers of devices with wide folded wrinkles, we find a distinct anisotropy in their electrical resistivity, consistent with our transport simulations. These results highlight the coupling between morphology and electronic properties, which has important practical implications for large-scale high-speed graphene electronics.

cond-mat.mes-hall↗

arXiv2011

Multilayer epitaxial graphene formed by pyrolysis of polycrystalline silicon-carbide grown on c-plane sapphire substrates

We use ultra-high vacuum chemical vapor deposition to grow polycrystalline silicon carbide (SiC) on c-plane sapphire wafers which are then annealed between 1250 and 1450°C in vacuum to create epitaxial multilayer graphene (MLG). Despite the surface roughness and small domain size of the polycrystalline SiC, a conformal MLG film is formed. By planarizing the SiC prior to graphene growth, a reduction of the Raman defect band is observed in the final MLG. The graphene formed on polished SiC films also demonstrates significantly more ordered layer-by-layer growth and increased carrier mobility for the same carrier density as the non-polished samples.

cond-mat.mtrl-sci↗

arXiv2010

Wafer-scale Epitaxial Graphene Growth on the Si-face of Hexagonal SiC (0001) for High Frequency Transistors

Up to two layers of epitaxial graphene have been grown on the Si-face of two-inch SiC wafers exhibiting room-temperature Hall mobilities up to 1800 cm^2/Vs, measured from ungated, large, 160 micron x 200 micron Hall bars, and up to 4000 cm^2/Vs, from top-gated, small, 1 micron x 1.5 micron Hall bars. The growth process involved a combination of a cleaning step of the SiC in a Si-containing gas, followed by an annealing step in Argon for epitaxial graphene formation. The structure and morphology of this graphene has been characterized using AFM, HRTEM, and Raman spectroscopy. Furthermore, top-gated radio frequency field effect transistors (RF-FETs) with a peak cutoff frequency fT of 100 GHz for a gate length of 240 nm were fabricated using epitaxial graphene grown on the Si face of SiC that exhibited Hall mobilities up to 1450 cm^2/Vs from ungated Hall bars and 1575 cm^2/Vs from top-gated ones. This is by far the highest cut-off frequency measured from any kind of graphene.

cond-mat.mtrl-sci↗

arXiv2003

Photon-assisted Fano Resonance and Corresponding Shot-Noise in a Quantum Dot

We have studied the Fano resonance in photon-assisted transport in a quantum dot and calculated both the coherent current and spectral density of shot noise. It is predicted, for the first time, that the shape of Fano profile will also appear in satellite peaks. It is found that the variations of Fano profiles with the strengths of nonresonant transmissions are not synchronous in absorption and emission sidebands. The effect of interference on photon-assisted pumped current has been also investigated. We further predict the current and spectral density of shot noise as a function of the phase, which exhibits an intrinsic property of resonant and nonresonant channels in the structures.

cond-mat.mes-hall↗