arXiv subjects

Jia Wang

Publications and source records attributed to Jia Wang.

At least 199 records · Page 11Linked to original sources

arXiv2020

KNN-enhanced Deep Learning Against Noisy Labels

Supervised learning on Deep Neural Networks (DNNs) is data hungry. Optimizing performance of DNN in the presence of noisy labels has become of paramount importance since collecting a large dataset will usually bring in noisy labels. Inspired by the robustness of K-Nearest Neighbors (KNN) against data noise, in this work, we propose to apply deep KNN for label cleanup. Our approach leverages DNNs for feature extraction and KNN for ground-truth label inference. We iteratively train the neural network and update labels to simultaneously proceed towards higher label recovery rate and better classification performance. Experiment results show that under the same setting, our approach outperforms existing label correction methods and achieves better accuracy on multiple datasets, e.g.,76.78% on Clothing1M dataset.

cs.CV↗

arXiv2020

A Learning-based Optimization Algorithm:Image Registration Optimizer Network

Remote sensing image registration is valuable for image-based navigation system despite posing many challenges. As the search space of registration is usually non-convex, the optimization algorithm, which aims to search the best transformation parameters, is a challenging step. Conventional optimization algorithms can hardly reconcile the contradiction of simultaneous rapid convergence and the global optimization. In this paper, a novel learning-based optimization algorithm named Image Registration Optimizer Network (IRON) is proposed, which can predict the global optimum after single iteration. The IRON is trained by a 3D tensor (9x9x9), which consists of similar metric values. The elements of the 3D tensor correspond to the 9x9x9 neighbors of the initial parameters in the search space. Then, the tensor's label is a vector that points to the global optimal parameters from the initial parameters. Because of the special architecture, the IRON could predict the global optimum directly for any initialization. The experimental results demonstrate that the proposed algorithm performs better than other classical optimization algorithms as it has higher accuracy, lower root of mean square error (RMSE), and more efficiency. Our IRON codes are available for further study.https://www.github.com/jaxwangkd04/IRON

cs.CV↗

arXiv2020

KeyGuard: Using Selective Encryption to Mitigate Keylogging in Third-Party IME

As mobile devices become ubiquitous, people around the world have enjoyed the convenience they have brought to our lives. At the same time, the increasing security threats that rise from using mobile devices not only have caught attention from cyber security agencies but also have become a valid concern for mobile users. Keylogging is one of the mobile security threats caused by using insecure third-party IME (input method editor) applications. Keylogging, as the name suggests, keeps track of user\rq s key events performed on the device and stores all the events in a log. The log could include highly sensitive data such as credit card number, social security number, and passwords. This paper presents a novel solution by intercepting the keystroke events triggered by a user and encrypting them before sending them to the third-party IME, making the third-party IME unable to log what the users actually entered on the screen. Input will be decrypted when showing on text view on the underlying app. This solution addresses the fundamental reason why an IME may leak sensitive information since an IME will no longer have access to the user\rq s actual sensitive information, which will greatly reduce the chance of leaking sensitive information by using a third-party IME while maintaining the functionalities of the third-party IME at the same time.

cs.CR↗

arXiv2020

Ultradilute self-bound quantum droplets in Bose-Bose mixtures at finite temperature

We theoretically investigate the finite-temperature structure and collective excitations of a self-bound ultradilute Bose droplet in a flat space realized in a binary Bose mixture with attractive inter-species interactions on the verge of mean-field collapse. As the droplet formation relies critically on the repulsive force provided by Lee-Huang-Yang quantum fluctuations, which can be easily compensated by thermal fluctuations, we find a significant temperature effect in the density distribution and collective excitation spectrum of the Bose droplet. A finite-temperature phase diagram as a function of the number of particles is determined. We show that the critical number of particles at the droplet-to-gas transition increases dramatically with increasing temperature. Towards the bulk threshold temperature for thermally destabilizing an infinitely large droplet, we find that the excitation-forbidden, self-evaporation region in the excitation spectrum, predicted earlier by Petrov using a zero-temperature theory, shrinks and eventually disappears. All the collective excitations, including both surface modes and compressional bulk modes, become softened at the droplet-to-gas transition. The predicted temperature effects of a self-bound Bose droplet in this work could be difficult to measure experimentally due to the lack of efficient thermometry at low temperatures. However, these effects may already present in the current cold-atom experiments.

cond-mat.quant-gas↗

arXiv2020

Generative Model without Prior Distribution Matching

Variational Autoencoder (VAE) and its variations are classic generative models by learning a low-dimensional latent representation to satisfy some prior distribution (e.g., Gaussian distribution). Their advantages over GAN are that they can simultaneously generate high dimensional data and learn latent representations to reconstruct the inputs. However, it has been observed that a trade-off exists between reconstruction and generation since matching prior distribution may destroy the geometric structure of data manifold. To mitigate this problem, we propose to let the prior match the embedding distribution rather than imposing the latent variables to fit the prior. The embedding distribution is trained using a simple regularized autoencoder architecture which preserves the geometric structure to the maximum. Then an adversarial strategy is employed to achieve a latent mapping. We provide both theoretical and experimental support for the effectiveness of our method, which alleviates the contradiction between topological properties' preserving of data manifold and distribution matching in latent space.

cs.CV↗

arXiv2020

Spatial Separation of the Conformers of Methyl Vinyl Ketone

Methyl vinyl ketone (C$_4$H$_6$O) is a volatile, labile organic compound of importance in atmospheric chemistry. We prepared a molecular beam of methyl vinyl ketone with a rotational temperature of 1.2(2)~K and demonstrated the spatial separation of the \emph{s-cis} and \emph{s-trans} conformers of methyl vinyl ketone using the electrostatic deflector. The resulting sample density was $1.5(2)\times10^{8}~\text{cm}^{-3}$ for the direct beam in the laser ionization region. These conformer-selected methyl vinyl ketone samples are well suited for conformer-specific chemical reactivity studies such as in Diels-Alder cycloaddition reactions.

physics.chem-ph↗

arXiv2020

Microscopic pairing theory of a binary Bose mixture with interspecies attractions: bosonic BEC-BCS crossover and ultradilute low-dimensional quantum droplets

Ultradilute quantum droplets are intriguing new state of matter, in which the attractive mean-field force can be balanced by the repulsive force from quantum fluctuations to avoid collapse. Here, we present a microscopic theory of ultradilute quantum droplets in three-, one- and two-dimensional two-component Bose-Bose mixtures, by generalizing the conventional Bogoliubov theory to include the bosonic pairing arising from the interspecies attraction. Our pairing theory is fully equivalent to a variational approach and hence gives an upper bound for the energy of quantum droplets. In three dimensions, we predict the existence of a strongly interacting Bose droplet at the crossover from Bose-Einstein condensates (BEC) to Bardeen--Cooper--Schrieffer (BCS) superfluids and map out the bosonic BEC-BCS crossover phase diagram. In one dimension, we find that the energy of the one-dimensional Bose droplet calculated by the pairing theory is in an excellent agreement with the latest diffusion Monte Carlo simulation {[}Phys. Rev. Lett. \textbf{122}, 105302 (2019){]}, for nearly all the interaction strengths at which quantum droplets exist. In two dimensions, we show that Bose droplets disappear and may turn into a soliton-like many-body bound state, when the interspecies attraction exceeds a critical value. Below the threshold, the pairing theory predicts more or less the same results as the Bogoliubov theory derived by Petrov and Astrakharchik {[}Phys. Rev. Lett. \textbf{117}, 100401 (2016){]}. The predicted energies from both theories are higher than the diffusion Monte Carlo results, due to the weak interspecies attraction and the increasingly important role played by the beyond-Bogoliubov-approximation effect in two dimensions.

cond-mat.quant-gas↗

arXiv2020

Uniform Interpolation Constrained Geodesic Learning on Data Manifold

In this paper, we propose a method to learn a minimizing geodesic within a data manifold. Along the learned geodesic, our method can generate high-quality interpolations between two given data samples. Specifically, we use an autoencoder network to map data samples into latent space and perform interpolation via an interpolation network. We add prior geometric information to regularize our autoencoder for the convexity of representations so that for any given interpolation approach, the generated interpolations remain within the distribution of the data manifold. Before the learning of a geodesic, a proper Riemannianmetric should be defined. Therefore, we induce a Riemannian metric by the canonical metric in the Euclidean space which the data manifold is isometrically immersed in. Based on this defined Riemannian metric, we introduce a constant speed loss and a minimizing geodesic loss to regularize the interpolation network to generate uniform interpolation along the learned geodesic on the manifold. We provide a theoretical analysis of our model and use image translation as an example to demonstrate the effectiveness of our method.

cs.CV↗

arXiv2020

Thermal destabilization of self-bound ultradilute quantum droplets

We theoretically investigate the temperature effect in a Bose-Bose mixture with attractive inter-species interactions, in the regime where a self-bound ultradilute quantum droplet forms due to the subtle balance between the attractive mean-field force and the repulsive force provided by Lee-Huang-Yang quantum fluctuations. We find that in contrast to quantum fluctuations, thermal fluctuations destabilize the droplet state and completely destroy it above a threshold temperature. We show that the threshold temperature is determined by the intra-species interaction energy. For a three-dimensional Bose-Bose mixture, the threshold temperature is less than one-tenth of the Bose-Einstein condensation temperature under the typical experimental conditions. With increasing temperature, the droplet's equilibrium density gradually decreases and can be reduced by several tens of percent upon reaching the threshold temperature. We also consider a one-dimensional quantum droplet and find a similar destabilization effect due to thermal fluctuations. The threshold temperature in one dimension is roughly set by the binding energy of the inter-species dimer. The pronounced thermal instability of a self-bound quantum droplet predicted in our work could be examined in future experiments, by measuring the temperature dependence of its central density and observing its sudden disappearance at the threshold temperature.

cond-mat.quant-gas↗

arXiv2020

Alleviation of Gradient Exploding in GANs: Fake Can Be Real

In order to alleviate the notorious mode collapse phenomenon in generative adversarial networks (GANs), we propose a novel training method of GANs in which certain fake samples are considered as real ones during the training process. This strategy can reduce the gradient value that generator receives in the region where gradient exploding happens. We show the process of an unbalanced generation and a vicious circle issue resulted from gradient exploding in practical training, which explains the instability of GANs. We also theoretically prove that gradient exploding can be alleviated by penalizing the difference between discriminator outputs and fake-as-real consideration for very close real and fake samples. Accordingly, Fake-As-Real GAN (FARGAN) is proposed with a more stable training process and a more faithful generated distribution. Experiments on different datasets verify our theoretical analysis.

cs.LG↗

arXiv2020

Spatial separation of 2-propanol monomer and its ionization-fragmentation pathways

The spatial separation of 2-propanol monomer from its clusters in a molecular beam by an electrostatic deflector was demonstrated. Samples of 2-propanol monomer with a purity of 90 % and a beam density of $7\times10^6~\text{cm}^{-3}$ were obtained. These samples were utilized to study the femtosecond-laser-induced strong-field multi-photon ionization and fragmentation of 2-propanol using non-resonant 800 nm light with peak intensities of $3-7\times10^{13}~W/cm^{2}$.

physics.chem-ph↗

arXiv2019

Many-body localization in XY spin chains with long-range interactions: An exact diagonalization study

We investigate the transition from the many-body localized phase to the ergodic thermalized phase at an infinite temperature in an $XY$ spin chain with $L$ spins, which experiences power-law decaying interactions in the form of $V_{ij}\propto1/\left|i-j\right|^α$ ($i,j=1,\cdots,L$) and a random transverse field. By performing large-scale exact diagonalization for the chain size up to $L=18$, we systematically analyze the energy gap statistics, half-chain entanglement entropy, and uncertainty of the entanglement entropy of the system at different interaction exponents $α$. The finite-size critical scaling allows us to determine the critical disorder strength $W_{c}$ and critical exponent $ν$ at the many-body localization phase transition, as a function of the interaction exponent $α$ in the limit $L\rightarrow\infty$. We find that both $W_{c}$ and $ν$ diverge when $α$ decreases to a critical power $α_{c}\simeq1.16\pm0.17$, indicating the absence of many-body localization for $α<α_{c}$. Our result is useful to resolve the contradiction on the critical power found in two previous studies, $α_{c}=3/2$ from scaling argument in Phys. Rev. B \textbf{92}, 104428 (2015) and $α_{c}\approx1$ from quantum dynamics simulation in Phys. Rev. A \textbf{99}, 033610 (2019).

cond-mat.quant-gas↗

arXiv2019

Efficient training and design of photonic neural network through neuroevolution

Recently, optical neural networks (ONNs) integrated in photonic chips has received extensive attention because they are expected to implement the same pattern recognition tasks in the electronic platforms with high efficiency and low power consumption. However, the current lack of various learning algorithms to train the ONNs obstructs their further development. In this article, we propose a novel learning strategy based on neuroevolution to design and train the ONNs. Two typical neuroevolution algorithms are used to determine the hyper-parameters of the ONNs and to optimize the weights (phase shifters) in the connections. In order to demonstrate the effectiveness of the training algorithms, the trained ONNs are applied in the classification tasks for iris plants dataset, wine recognition dataset and modulation formats recognition. The calculated results exhibit that the training algorithms based on neuroevolution are competitive with other traditional learning algorithms on both accuracy and stability. Compared with previous works, we introduce an efficient training method for the ONNs and demonstrate their broad application prospects in pattern recognition, reinforcement learning and so on.

cs.NE↗

arXiv2019

Roton-induced Bose polaron in the presence of synthetic spin-orbit coupling

We theoretically investigate the quasiparticle (polaron) properties of an impurity immersing in a Bose-Einstein condensate with equal Rashba and Dresselhaus spin-orbit coupling at zero temperature. In the presence of spin-orbit coupling, all bosons can condense into a single plane-wave state with finite momentum, and the corresponding excitation spectrum shows an intriguing roton minimum. We find that the polaron properties are strongly modified by this roton minimum, where the ground state of attractive polaron acquires a nonzero momentum and anisotropic effective mass. Across the resonance of the interaction between impurity and atoms, the polaron evolves into a tight-binding dimer. We show that the evolution is not smooth when the roton structure of the condensate becomes apparent, and a first-order phase transition from a phonon-induced polaron to a roton-induced polaron is observed at a critical interaction strength.

cond-mat.quant-gas↗

arXiv2018

Efficient spectrum prediction and inverse design for plasmonic waveguide systems based on artificial neural networks

In this article, we propose a novel approach to achieve spectrum prediction, parameter fitting, inverse design and performance optimization for the plasmonic waveguide coupled with cavities structure (PWCCS) based on artificial neural networks (ANNs). The Fano resonance and plasmon induced transparency effect originated from the PWCCS have been selected as illustrations to verify the effectiveness of ANNs. We use the genetic algorithm to design the network architecture and select the hyper-parameters for ANNs. Once ANNs are trained by using a small sampling of the data generated by Monte Carlo method, the transmission spectrums predicted by the ANNs are quite approximate to the simulated results. The physical mechanisms behind the phenomena are discussed theoretically, and the uncertain parameters in the theoretical models are fitted by utilizing the trained ANNs. More importantly, our results demonstrate that this model-driven method not only realizes the inverse design of the PWCCS with high precision but also optimizes some critical performance metrics for transmission spectrum. Compared with previous works, we construct a novel model-driven analysis method for the PWCCS which are expected to have significant applications in the device design, performance optimization, variability analysis, defect detection, theoretical modeling, optical interconnects and so on.

physics.optics↗

arXiv2018

Understanding VAEs in Fisher-Shannon Plane

In information theory, Fisher information and Shannon information (entropy) are respectively used to quantify the uncertainty associated with the distribution modeling and the uncertainty in specifying the outcome of given variables. These two quantities are complementary and are jointly applied to information behavior analysis in most cases. The uncertainty property in information asserts a fundamental trade-off between Fisher information and Shannon information, which enlightens us the relationship between the encoder and the decoder in variational auto-encoders (VAEs). In this paper, we investigate VAEs in the Fisher-Shannon plane and demonstrate that the representation learning and the log-likelihood estimation are intrinsically related to these two information quantities. Through extensive qualitative and quantitative experiments, we provide with a better comprehension of VAEs in tasks such as high-resolution reconstruction, and representation learning in the perspective of Fisher information and Shannon information. We further propose a variant of VAEs, termed as Fisher auto-encoder (FAE), for practical needs to balance Fisher information and Shannon information. Our experimental results have demonstrated its promise in improving the reconstruction accuracy and avoiding the non-informative latent code as occurred in previous works.

stat.ML↗

arXiv2018

Quantum fluctuations of a resonantly interacting $p$-wave Fermi superfluid in two dimensions

Using the Gaussian pair fluctuation theory, we investigate quantum fluctuations of a strongly interacting two-dimensional chiral \textit{p}-wave Fermi superfluid at the transition from a Bose-Einstein condensate (BEC) to a topologically non-trivial Bardeen-Cooper-Schrieffer (BCS) superfluid. Near the topological phase transition at zero chemical potential, $μ=0$, we observe that quantum fluctuations strongly renormalize the zero-temperature equations of state, sound velocity, pair-breaking velocity, and Berezinskii-Kosterlitz-Thouless (BKT) critical temperature of the Fermi superfluid, all of which can be non-analytic functions of the interaction strength. The indication of non-analyticity is particularly evident in the BKT critical temperature, which also exhibits a pronounced peak near the topological phase transition. Across the transition and towards the BEC limit we find that the system quickly becomes a trivial interacting Bose liquid, whose properties are less dependent on the interparticle interaction. The qualitative behavior of composite bosons in the BEC limit remains to be understood.

cond-mat.quant-gas↗

arXiv2018

Quantum depletion and superfluid density of a supersolid in Raman spin-orbit coupled Bose gases

We theoretically investigate a three-dimensional weakly interacting Bose gas with one-dimensional Raman-type spin-orbit coupling at zero temperature. By employing an improved ansatz, including high-order harmonics in the stripe phase, we show that the critical transition from the stripe to the plane-wave phases is shifted to a relatively larger Rabi frequency compared to the prediction by previous work [Li $\textit{et al.}$, Phys. Rev. Lett. $\textbf{108}$, 225301 (2012)] using a first-order stripe ansatz. We also determine the quantum depletion and superfluid density over a large range of Rabi frequency in different phases. The depletion exhibits an intriguing behavior with a discontinuous jump at the transition between the stripe and plane-wave phases, and a maximum at the transition between the plane-wave and zero-momentum phases. The superfluid density is derived through a phase-twist method. In the plane-wave and zero-momentum phases, it is significantly suppressed along the spin-orbit-coupling direction and vanishes at the transition, consistent with a recent work [Zhang $\textit{et al.}$, Phys. Rev. A $\textbf{94}$, 033635 (2016)], while in the stripe phase, it smoothly decreases with increasing Rabi frequency. Our predictions would be useful for further theoretical and experimental studies of the exotic supersolid stripe phase.

cond-mat.quant-gas↗