Searcharxiv⌕ Search

arXiv subjects

Wei Zhu

Publications and source records attributed to Wei Zhu.

At least 253 records · Page 14Linked to original sources
arXiv2020

Adversarial Defense via Data Dependent Activation Function and Total Variation Minimization

We improve the robustness of Deep Neural Net (DNN) to adversarial attacks by using an interpolating function as the output activation. This data-dependent activation remarkably improves both the generalization and robustness of DNN. In the CIFAR10 benchmark, we raise the robust accuracy of the adversarially trained ResNet20 from $\sim 46\%$ to $\sim 69\%$ under the state-of-the-art Iterative Fast Gradient Sign Method (IFGSM) based adversarial attack. When we combine this data-dependent activation with total variation minimization on adversarial images and training data augmentation, we achieve an improvement in robust accuracy by 38.9$\%$ for ResNet56 under the strongest IFGSM attack. Furthermore, We provide an intuitive explanation of our defense by analyzing the geometry of the feature space.

cs.LG↗
arXiv2020

Alleviating the Incompatibility between Cross Entropy Loss and Episode Training for Few-shot Skin Disease Classification

Skin disease classification from images is crucial to dermatological diagnosis. However, identifying skin lesions involves a variety of aspects in terms of size, color, shape, and texture. To make matters worse, many categories only contain very few samples, posing great challenges to conventional machine learning algorithms and even human experts. Inspired by the recent success of Few-Shot Learning (FSL) in natural image classification, we propose to apply FSL to skin disease identification to address the extreme scarcity of training sample problem. However, directly applying FSL to this task does not work well in practice, and we find that the problem can be largely attributed to the incompatibility between Cross Entropy (CE) and episode training, which are both commonly used in FSL. Based on a detailed analysis, we propose the Query-Relative (QR) loss, which proves superior to CE under episode training and is closely related to recently proposed mutual information estimation. Moreover, we further strengthen the proposed QR loss with a novel adaptive hard margin strategy. Comprehensive experiments validate the effectiveness of the proposed FSL scheme and the possibility to diagnosis rare skin disease with a few labeled samples.

eess.IV↗
arXiv2020

A robust fiber-based quantum thermometer coupled with nitrogen-vacancy centers

The nitrogen-vacancy center in diamond has been broadly applied in quantum sensing since it is sensitive to different physical quantities. Meanwhile, it is difficult to isolate disturbances from unwanted physical quantities in practical applications. Here, we present a robust fiber-based quantum thermometer which can significantly isolate the magnetic field noise and microwave power shift. With a frequency modulation scheme, we realize the temperature measurement by detecting the variation of the sharp-dip in the zero-field optically detected magnetic resonance spectrum in a high-density nitrogen-vacancy ensemble. Thanks to its simplicity and compatibility in implementation and robustness in the isolation of magnetic and microwave noise, this quantum thermometer is then applied to the surface temperature imaging of an electronic chip with a sensitivity of $18$ $\rm{mK}/\sqrt{\rm{Hz}}$. It paves the way to high sensitive temperature measurement in ambiguous environments.

physics.app-ph↗
arXiv2020

Weighted Aggregating Stochastic Gradient Descent for Parallel Deep Learning

This paper investigates the stochastic optimization problem with a focus on developing scalable parallel algorithms for deep learning tasks. Our solution involves a reformation of the objective function for stochastic optimization in neural network models, along with a novel parallel strategy, coined weighted aggregating stochastic gradient descent (WASGD). Following a theoretical analysis on the characteristics of the new objective function, WASGD introduces a decentralized weighted aggregating scheme based on the performance of local workers. Without any center variable, the new method automatically assesses the importance of local workers and accepts them according to their contributions. Furthermore, we have developed an enhanced version of the method, WASGD+, by (1) considering a designed sample order and (2) applying a more advanced weight evaluating function. To validate the new method, we benchmark our schemes against several popular algorithms including the state-of-the-art techniques (e.g., elastic averaging SGD) in training deep neural networks for classification tasks. Comprehensive experiments have been conducted on four classic datasets, including the CIFAR-100, CIFAR-10, Fashion-MNIST, and MNIST. The subsequent results suggest the superiority of the WASGD scheme in accelerating the training of deep architecture. Better still, the enhanced version, WASGD+, has been shown to be a significant improvement over its basic version.

cs.LG↗
arXiv2020

Unifying Specialist Image Embedding into Universal Image Embedding

Deep image embedding provides a way to measure the semantic similarity of two images. It plays a central role in many applications such as image search, face verification, and zero-shot learning. It is desirable to have a universal deep embedding model applicable to various domains of images. However, existing methods mainly rely on training specialist embedding models each of which is applicable to images from a single domain. In this paper, we study an important but unexplored task: how to train a single universal image embedding model to match the performance of several specialists on each specialist's domain. Simply fusing the training data from multiple domains cannot solve this problem because some domains become overfitted sooner when trained together using existing methods. Therefore, we propose to distill the knowledge in multiple specialists into a universal embedding to solve this problem. In contrast to existing embedding distillation methods that distill the absolute distances between images, we transform the absolute distances between images into a probabilistic distribution and minimize the KL-divergence between the distributions of the specialists and the universal embedding. Using several public datasets, we validate that our proposed method accomplishes the goal of universal image embedding.

cs.CV↗
arXiv2020

Quantifying the performance of multi-pulse quantum sensing

The quality of a quantum operation determines the performance of quantum information processing, such as the sensitivity of quantum sensing. Different from the fidelity of quantum operation in quantum computation, we present an effective function to evaluate the performance and diagnose the imperfection of operations in multi-pulse based quantum sensing. The evaluation function directly links the realistic sensitivity with intrinsic sensitivity in a simple way. Moreover, guided by this evaluation function, we optimize a composite pulse sequence for high sensitivity nitrogen-vacancycenter based magnetometry against spectrum inhomogeneities and control errors to improve the signal-to-noise ratio of nanoscale nuclear magnetic resonance by 1 order of magnitude. It marks an important step towards quantitative quantum sensing with imperfect quantum control in practical applications.

quant-ph↗
arXiv2020

Improved Bethe-Heitler formula

The Bethe-Heitler formula describes bremsstrahlung and it's a typical and important example in quantum electromagnetic dynamics (QED). This formula is widely applied in many branches of physics and astrophysics. We find that the integrated bremsstrahlung cross section at the static approximation and high energy limit has an unexpected big increment, which is missed by previous bremsstrahlung theory. This anomalous effect also exists in electron-positron pair creation. We derive the relating formulas and point out that electromagnetic cascades at the top of atmosphere can test this effect.

hep-ph↗
arXiv2020

On the patterns observed in Kepler multi-planet systems

Recent studies claimed that planets around the same star have similar sizes and masses and regular spacings, and that planet pairs usually show ordered sizes such that the outer planet is usually the larger one. Here I show that these patterns can be largely explained by detection biases. The \emph{Kepler} planet detections are set by the transit signal-to-noise ratio (S/N). For different stellar properties and orbital period values, the same S/N corresponds to different planetary sizes. This variation in the detection threshold naturally leads to apparent correlations in planet sizes and the observed size ordering. The apparently correlated spacings, measured in period ratios, between adjacent planet pairs in systems with at least three detected planets are partially due to the arbitrary upper limit that the earlier study imposed on the period ratio, and partially due to the varying stability threshold for different planets. After these detection biases are taken into account, we do not find strong evidence for the so-called "intra-system uniformity" or the size ordering effect. Instead, the physical properties of \emph{Kepler} planets are largely independent of the properties of their siblings and the parent star. It is likely that the dynamical evolution has erased the memory of \emph{Kepler} planets about their initial formation conditions. In other words, it will be difficult to infer the initial conditions from the observed properties and the architecture of \emph{Kepler} planets.

astro-ph.EP↗
arXiv2019

KMT-2018-BLG-0029Lb: A Very Low Mass-Ratio Spitzer Microlens Planet

At $q=1.81\pm 0.20 \times 10^{-5}$, KMT-2018-BLG-0029Lb has the lowest planet-host mass ratio $q$ of any microlensing planet to date by more than a factor of two. Hence, it is the first planet that probes below the apparent "pile-up" at $q=5$--10 $\times 10^{-5}$. The event was observed by {\it Spitzer}, yielding a microlens-parallax $π_{\rm E}$ measurement. Combined with a measurement of the Einstein radius $θ_{\rm E}$ from finite-source effects during the caustic crossings, these measurements imply masses of the host $M_{\rm host}=1.14^{+0.10}_{-0.12}\, M_\odot$ and planet $M_{\rm planet} = 7.59^{+0.75}_{-0.69}\,M_\oplus$, system distance $D_L = 3.38^{+0.22}_{-0.26}\,\,{\rm kpc}$ and projected separation $a_\perp = 4.27^{+0.21}_{-0.23}\,{\rm au}$. The blended light, which is substantially brighter than the microlensed source, is plausibly due to the lens and could be observed at high resolution immediately.

astro-ph.EP↗
arXiv2019

KMT-2016-BLG-1836Lb: A Super-Jovian Planet From A High-Cadence Microlensing Field

We report the discovery of a super-Jovian planet in the microlensing event KMT-2016-BLG-1836, which was found by the Korea Microlensing Telescope Network's high-cadence observations (Γ~ 4~{hr}^{-1}). The planet-host mass ratio q ~ 0.004. A Bayesian analysis indicates that the planetary system is composed of a super-Jovian M_{planet} = 2.2_{-1.1}^{+1.9} M_{J} planet orbiting an M or K dwarf M_{\rm host} = 0.49_{-0.25}^{+0.38} M_{Sun}, at a distance of D_{L} = 7.1_{-2.4}^{+0.8} kpc. The projected planet-host separation is 3.5^{+1.1}_{-0.9} AU, implying that the planet is located beyond the snowline of the host star. Future high-resolution images can potentially strongly constrain the lens brightness and thus the mass and distance of the planetary system. Without considering detailed detection efficiency, selection or publication biases, we find a potential "mass ratio desert" at -3.7 \lesssim \log q \lesssim -3.0 for the 31 published KMTNet planets.

astro-ph.EP↗
arXiv2019

A simplified connection between constituent quark and parton

We propose a simple way to connect Gell-Mann constituent quark model and Feynman parton model for the nucleon. Thus, we can dynamically understand a large amount of data for high energy hadronic processes starting from the first QCD principle.

hep-ph↗
arXiv2019

Efficient Identifying the Orientation of Single NV Centers in Diamond and Using them to Detect Near Field Microwave

Arrays of NV centers in the diamond have the potential in the fields of chip-scale quantum information processing and nanoscale quantum sensing. However, determining their orientations one by one is resource intensive and time consuming. Here, in this paper, by combining scanning confocal fluorescence images and optical detected magnetic resonance, we realized a method of identifying single NV centers with the same orientation, which is practicable and high efficiency. In the proof of principle experiment, five single NV centers with the same orientation in a NV center array were identified. After that, using the five single NV centers, microwave near field generated by a 20 μm-diameter Cu antenna was also measured by reading the fluourescence intensity change and Rabi frequency at different microwave source power. The gradient of near field microwave at sub-microscale can be resoluted by using arry of NV centers in our work. This work promotes the quantum sensing using arrays of NV centers.

physics.app-ph↗
arXiv2019

Exploiting Color Name Space for Salient Object Detection

In this paper, we will investigate the contribution of color names for the task of salient object detection. An input image is first converted to color name space, which is consisted of 11 probabilistic channels. By exploiting a surroundedness cue, we obtain a saliency map through a linear combination of a set of sequential attention maps. To overcome the limitation of only using the surroundedness cue, two global cues with respect to color names are invoked to guide the computation of a weighted saliency map. Finally, we integrate the above two saliency maps into a unified framework to generate the final result. In addition, an improved post-processing procedure is introduced to effectively suppress image backgrounds while uniformly highlight salient objects. Experimental results show that the proposed model produces more accurate saliency maps and performs well against twenty-one saliency models in terms of three evaluation metrics on three public data sets.

cs.CV↗
arXiv2019

Spitzer + VLTI-GRAVITY Measure the Lens Mass of a Nearby Microlensing Event

We report the lens mass and distance measurements of the nearby microlensing event TCP J05074264+2447555. We measure the microlens parallax vector $π_{\rm E}$ using Spitzer and ground-based light curves with constraints on the direction of lens-source relative proper motion derived from Very Large Telescope Interferometer (VLTI) GRAVITY observations. Combining this $π_{\rm E}$ determination with the angular Einstein radius $θ_{\rm E}$ measured by VLTI GRAVITY observations, we find that the lens is a star with mass $M_{\rm L} = 0.495 \pm 0.063~M_{\odot}$ at a distance $D_{\rm L} = 429 \pm 21~{\rm pc}$. We find that the blended light basically all comes from the lens. The lens-source proper motion is $μ_{\rm rel,hel} = 26.55 \pm 0.36~{\rm mas\,yr^{-1}}$, so with currently available adaptive-optics (AO) instruments, the lens and source can be resolved in 2021. This is the first microlensing event whose lens mass is unambiguously measured by interferometry + satellite parallax observations, which opens a new window for mass measurements of isolated objects such as stellar-mass black holes.

astro-ph.IM↗
arXiv2019

OGLE-2015-BLG-1771Lb: A Microlens Planet Orbiting an Ultracool Dwarf?

We report the discovery and the analysis of the short (tE < 5 days) planetary microlensing event, OGLE-2015-BLG-1771. The event was discovered by the Optical Gravitational Lensing Experiment (OGLE), and the planetary anomaly (at I ~ 19) was captured by The Korea Microlensing Telescope Network (KMTNet). The event has three surviving planetary models that explain the observed light curves, with planet-host mass ratio q \~ 5.4 * 10^{-3}, 4.5 * 10^{-3} and 4.5 * 10^{-2}, respectively. The first model is the best-fit model, while the second model is disfavored by Δχ^2 ~ 3. The last model is strongly disfavored by Δχ^2 ~ 15 but not ruled out. A Bayesian analysis using a Galactic model indicates that the first two models are probably composed of a Saturn-mass planet orbiting a late M dwarf, while the third one could consist of a super-Jovian planet and a mid-mass brown dwarf. The source-lens relative proper motion is mu_rel ~ 9 mas/yr, so the source and lens could be resolved by current adaptive-optics (AO) instruments in 2021 if the lens is luminous.

astro-ph.EP↗
arXiv2019

Spitzer Microlensing Program as a Probe for Globular Cluster Planets. Analysis of OGLE-2015-BLG-0448

The microlensing event OGLE-2015-BLG-0448 was observed by Spitzer and lay within the tidal radius of the globular cluster NGC 6558. The event had moderate magnification and was intensively observed, hence it had the potential to probe the distribution of planets in globular clusters. We measure the proper motion of NGC 6558 ($μ_{\rm cl}$(N,E) = (+0.36+-0.10, +1.42+-0.10) mas/yr) as well as the source and show that the lens is not a cluster member. Even though this particular event does not probe the distribution of planets in globular clusters, other potential cluster lens events can be verified using our methodology. Additionally, we find that microlens parallax measured using OGLE photometry is consistent with the value found based on the light curve displacement between Earth and Spitzer.

astro-ph.SR↗
arXiv2019

Large Effects of Particle Size Heterogeneity on Dynamic Saltation Threshold

Reliably predicting the geomorphology and climate of planetary bodies requires knowledge of the dynamic threshold wind shear velocity below which saltation transport ceases. Here we measure this threshold in a wind tunnel for four well-sorted and two poorly sorted sand beds by visual means and by a method that exploits a regime shift in the behavior of the surface roughness caused by momentum transfer from the wind to the saltating particles. For our poorly sorted sands, we find that these measurement methods yield different threshold values because, at the smaller visual threshold, relatively coarse particles do not participate in saltation. We further find that both methods yield threshold values that are much larger (60--250\%) for our poorly sorted sands than for our well-sorted sands with similar median particle diameter. In particular, even a rescaling of the dynamic saltation threshold based on the 90th percentile particle diameter rather than the median diameter cannot fully capture this difference, suggesting that relatively very coarse particles have a considerable control on the dynamic threshold. Similar findings were previously reported for water-driven sediment transport. Our findings have important implications for quantitative predictions of saltation transport-related geophysical processes, such as dust aerosol emission.

physics.geo-ph↗