arXiv subjects

Min Du

Publications and source records attributed to Min Du.

At least 55 records · Page 3Linked to original sources

arXiv2020

Learning Enhanced Resolution-wise features for Human Pose Estimation

Recently, multi-resolution networks (such as Hourglass, CPN, HRNet, etc.) have achieved significant performance on pose estimation by combining feature maps of various resolutions. In this paper, we propose a Resolution-wise Attention Module (RAM) and Gradual Pyramid Refinement (GPR), to learn enhanced resolution-wise feature maps for precise pose estimation. Specifically, RAM learns a group of weights to represent the different importance of feature maps across resolutions, and the GPR gradually merges every two feature maps from low to high resolutions to regress final human keypoint heatmaps. With the enhanced resolution-wise features learnt by CNN, we obtain more accurate human keypoint locations. The efficacies of our proposed methods are demonstrated on MS-COCO dataset, achieving state-of-the-art performance with average precision of 77.7 on COCO val2017 set and 77.0 on test-dev2017 set without using extra human keypoint training dataset.

cs.CV↗

arXiv2020

Ram pressure stripping of HI-rich galaxies infalling into massive clusters

We estimate the strength of ram pressure stripping (RPS) for HI-rich galaxies in X-ray detected clusters. We find that galaxies under stronger RPS tend to show more significantly reduced total HI mass and enhanced central SFR, compared to control galaxies in the field which have similar stellar mass, stellar surface density and integral star formation rate. Galaxies under strong or weak RPS account for around 40% of the HI-rich population at R200, and even beyond R200 in the most massive clusters. Our results imply the important role of RPS as a channel of environmental processing far before the galaxies reach the core region of clusters.

astro-ph.GA↗

arXiv2020

Barred Galaxies in the IllustrisTNG Simulation

Almost two-thirds of disk galaxies in the local universe host bars, which serve as important drivers of secular evolutionary processes. While cosmological simulations are powerful tools to study the formation and evolution of galaxies, they have often struggled to generate reasonable bar populations. We measure the fraction, size, and strength of bars in 3866 disk galaxies from the TNG100 run of the advanced cosmological simulation IllustrisTNG. Consistent with observations, about 55% of disk galaxies with stellar mass $M_* \approx 10^{10.6}\, M_\odot$ are barred, and the relation between bar size and total stellar mass is similar to that found in near-infrared surveys. However, the formation of bars is suppressed in galaxies with $M_* < 10^{10.6}\, M_\odot$, which may result from the difficulty TNG100 has in resolving short bars with radius <1.4 kpc. In contrast, up to 75% of massive disk galaxies with $M_* > 10^{10.6}\, M_\odot$ have bars, ~10%-20% higher than observed. TNG100 overproduces relatively short bars (radius ~1.4-3 kpc) with respect to the mass-bar size relation observed in near-infrared surveys. Tracing the progenitors of z=0 massive galaxies we find that the bar fraction increases from 25% to 63% between z=1 and 0. Instead if we select all disk galaxies during z=0-1 with a constant mass cut of $M_* \geqslant 10^{10.6}\, M_\odot$ we find that the bar fraction is a nearly constant 60%.

astro-ph.GA↗

arXiv2020

Kinematic decomposition of IllustrisTNG disk galaxies: morphology and relation with morphological structures

We recently developed an automated method, auto-GMM to decompose simulated galaxies. It extracts kinematic structures in an accurate, efficient, and unsupervised way. We use auto-GMM to study the stellar kinematic structures of disk galaxies from the TNG100 run of IllustrisTNG. We identify four to five structures that are commonly present among the diverse galaxy population. Structures having strong to moderate rotation are defined as cold and warm disks, respectively. Spheroidal structures dominated by random motions are classified as bulges or stellar halos, depending on how tightly bound they are. Disky bulges are structures that have moderate rotation but compact morphology. Across all disky galaxies and accounting for the stellar mass within 3 half-mass radii, the kinematic spheroidal structures, obtained by summing up stars of bulges and halos, contribute ~45% of the total stellar mass, while the disky structures constitute 55%. This study also provides important insights about the relationship between kinematically and morphologically derived galactic structures. Comparing the morphology of kinematic structures with that of traditional bulge+disk decomposition, we conclude: (1) the morphologically decomposed bulges are composite structures comprised of a slowly rotating bulge, an inner halo, and a disky bulge; (2) kinematically disky bulges, akin to what are commonly called pseudo bulges in observations, are compact disk-like components that have rotation similar to warm disks; (3) halos contribute almost 30% of the surface density of the outer part of morphological disks when viewed face-on; and (4) both cold and warm disks are often truncated in central regions.

astro-ph.GA↗

arXiv2020

Time-aware Gradient Attack on Dynamic Network Link Prediction

In network link prediction, it is possible to hide a target link from being predicted with a small perturbation on network structure. This observation may be exploited in many real world scenarios, for example, to preserve privacy, or to exploit financial security. There have been many recent studies to generate adversarial examples to mislead deep learning models on graph data. However, none of the previous work has considered the dynamic nature of real-world systems. In this work, we present the first study of adversarial attack on dynamic network link prediction (DNLP). The proposed attack method, namely time-aware gradient attack (TGA), utilizes the gradient information generated by deep dynamic network embedding (DDNE) across different snapshots to rewire a few links, so as to make DDNE fail to predict target links. We implement TGA in two ways: one is based on traversal search, namely TGA-Tra; and the other is simplified with greedy search for efficiency, namely TGA-Gre. We conduct comprehensive experiments which show the outstanding performance of TGA in attacking DNLP algorithms.

cs.SI↗

arXiv2020

The Formation History of Subhalos and the Evolution of Satellite Galaxies

Satellites constitute an important fraction of the overall galaxy population and are believed to form in dark matter subhalos. Here we use the cosmological hydrodynamic simulation TNG100 to investigate how the formation histories of subhalos affect the properties and evolution of their host galaxies. We use a scaled formation time ($a_{\rm nf}$) to characterize the mass assembly histories of the subhalos before they are accreted by massive host halos. We find that satellite galaxies in young subhalos (low $a_{\rm nf}$) are less massive and more gas rich, and have stronger star formation and a higher fraction of ex situ stellar mass than satellites in old subhalos (high $a_{\rm nf}$). Furthermore, these low $a_{\rm nf}$ satellites require longer timescales to be quenched as a population than the high $a_{\rm nf}$ counterparts. We find very different merger histories between satellites in fast accretion (FA, $a_{\rm nf}<1.3$) and slow accretion (SA, $a_{\rm nf}>1.3$) subhalos. For FA satellites, the galaxy merger frequency dramatically increases just after accretion, which enhances the star formation at accretion. While, for SA satellites, the mergers occur smoothly and continuously across the accretion time. Moreover, mergers with FA satellites happen mainly after accretion, while a contrary trend is found for SA satellites. Our results provide insight into the evolution and star formation quenching of the satellite population.

astro-ph.GA↗

arXiv2019

Free-riders in Federated Learning: Attacks and Defenses

Federated learning is a recently proposed paradigm that enables multiple clients to collaboratively train a joint model. It allows clients to train models locally, and leverages the parameter server to generate a global model by aggregating the locally submitted gradient updates at each round. Although the incentive model for federated learning has not been fully developed, it is supposed that participants are able to get rewards or the privilege to use the final global model, as a compensation for taking efforts to train the model. Therefore, a client who does not have any local data has the incentive to construct local gradient updates in order to deceive for rewards. In this paper, we are the first to propose the notion of free rider attacks, to explore possible ways that an attacker may construct gradient updates, without any local training data. Furthermore, we explore possible defenses that could detect the proposed attacks, and propose a new high dimensional detection method called STD-DAGMM, which particularly works well for anomaly detection of model parameters. We extend the attacks and defenses to consider more free riders as well as differential privacy, which sheds light on and calls for future research in this field.

cs.LG↗

arXiv2019

Robust Anomaly Detection and Backdoor Attack Detection Via Differential Privacy

Outlier detection and novelty detection are two important topics for anomaly detection. Suppose the majority of a dataset are drawn from a certain distribution, outlier detection and novelty detection both aim to detect data samples that do not fit the distribution. Outliers refer to data samples within this dataset, while novelties refer to new samples. In the meantime, backdoor poisoning attacks for machine learning models are achieved through injecting poisoning samples into the training dataset, which could be regarded as "outliers" that are intentionally added by attackers. Differential privacy has been proposed to avoid leaking any individual's information, when aggregated analysis is performed on a given dataset. It is typically achieved by adding random noise, either directly to the input dataset, or to intermediate results of the aggregation mechanism. In this paper, we demonstrate that applying differential privacy can improve the utility of outlier detection and novelty detection, with an extension to detect poisoning samples in backdoor attacks. We first present a theoretical analysis on how differential privacy helps with the detection, and then conduct extensive experiments to validate the effectiveness of differential privacy in improving outlier detection, novelty detection, and backdoor attack detection.

cs.LG↗

arXiv2019

A New Channel of Bulge Formation via The Destruction of Short Bars

Short (inner) bars of sub-kiloparsec radius have been hypothesized to be an important mechanism for driving gas inflows to small scales, thus feeding central black holes. Recent numerical simulations have shown that the growth of central black holes in galaxies can destroy short bars, when the black hole reaches a mass of $\sim 0.1\%$ of the total stellar mass of the galaxy. We study $N$-body simulations of galaxies with single and double bars to track the long-term evolution of the central stellar mass distribution. We find that the destruction of the short bar contributes significantly to the growth of the bulge. The final bulge mass is roughly equal to the sum of the masses of the initial pseudo bulge and short bar. The initially boxy/peanut-shaped bulge of Sérsic index $n\lesssim1$ is transformed into a more massive, compact structure that bears many similarities to a classical bulge, in terms of its morphology ($n \approx 2$), kinematics (dispersion-dominated, isotropic), and location on standard scaling relations (Kormendy relation, mass-size relation, and correlations between black hole mass and bulge stellar mass and velocity dispersion). Our proposed channel for forming classical bulges relies solely on the destruction of short bars without any reliance on mergers. We suggest that some of the less massive, less compact classical bulges were formed in this manner.

astro-ph.GA↗

arXiv2019

Stain Style Transfer using Transitive Adversarial Networks

Digitized pathological diagnosis has been in increasing demand recently. It is well known that color information is critical to the automatic and visual analysis of pathological slides. However, the color variations due to various factors not only have negative impact on pathologist's diagnosis, but also will reduce the robustness of the algorithms. The factors that cause the color differences are not only in the process of making the slices, but also in the process of digitization. Different strategies have been proposed to alleviate the color variations. Most of such techniques rely on collecting color statistics to perform color matching across images and highly dependent on a reference template slide. Since the pathological slides between hospitals are usually unpaired, these methods do not yield good matching results. In this work, we propose a novel network that we refer to as Transitive Adversarial Networks (TAN) to transfer the color information among slides from different hospitals or centers. It is not necessary for an expert to pick a representative reference slide in the proposed TAN method. We compare the proposed method with the state-of-the-art methods quantitatively and qualitatively. Compared with the state-of-the-art methods, our method yields an improvement of 0.87dB in terms of PSNR, demonstrating the effectiveness of the proposed TAN method in stain style transfer.

eess.IV↗

arXiv2019

Identifying kinematic structures in simulated galaxies using unsupervised machine learning

Galaxies host a wide array of internal stellar components, which need to be decomposed accurately in order to understand their formation and evolution. While significant progress has been made with recent integral-field spectroscopic surveys of nearby galaxies, much can be learned from analyzing the large sets of realistic galaxies now available through state-of-the-art hydrodynamical cosmological simulations. We present an unsupervised machine learning algorithm, named auto-GMM, based on Gaussian mixture models, to isolate intrinsic structures in simulated galaxies based on their kinematic phase space. For each galaxy, the number of Gaussian components allowed by the data is determined through a modified Bayesian information criterion. We test our method by applying it to prototype galaxies selected from the cosmological simulation IllustrisTNG. Our method can effectively decompose most galactic structures. The intrinsic structures of simulated galaxies can be inferred statistically by non-human supervised identification of galaxy structures. We successfully identify four kinds of intrinsic structures: cold disks, warm disks, bulges, and halos. Our method fails for barred galaxies because of the complex kinematics of particles moving on bar orbits.

astro-ph.GA↗

arXiv2019

TABOR: A Highly Accurate Approach to Inspecting and Restoring Trojan Backdoors in AI Systems

A trojan backdoor is a hidden pattern typically implanted in a deep neural network. It could be activated and thus forces that infected model behaving abnormally only when an input data sample with a particular trigger present is fed to that model. As such, given a deep neural network model and clean input samples, it is very challenging to inspect and determine the existence of a trojan backdoor. Recently, researchers design and develop several pioneering solutions to address this acute problem. They demonstrate the proposed techniques have a great potential in trojan detection. However, we show that none of these existing techniques completely address the problem. On the one hand, they mostly work under an unrealistic assumption (e.g. assuming availability of the contaminated training database). On the other hand, the proposed techniques cannot accurately detect the existence of trojan backdoors, nor restore high-fidelity trojan backdoor images, especially when the triggers pertaining to the trojan vary in size, shape and position. In this work, we propose TABOR, a new trojan detection technique. Conceptually, it formalizes a trojan detection task as a non-convex optimization problem, and the detection of a trojan backdoor as the task of resolving the optimization through an objective function. Different from the existing technique also modeling trojan detection as an optimization problem, TABOR designs a new objective function--under the guidance of explainable AI techniques as well as heuristics--that could guide optimization to identify a trojan backdoor in a more effective fashion. In addition, TABOR defines a new metric to measure the quality of a trojan backdoor identified. Using an anomaly detection method, we show the new metric could better facilitate TABOR to identify intentionally injected triggers in an infected model and filter out false alarms......

cs.CR↗

arXiv2019

The Formation of Compact Elliptical Galaxies in The Vicinity of A Massive Galaxy: The Role of Ram-pressure Confinement

Compact ellipticals (cEs) are outliers from the scaling relations of early-type galaxies, particularly the mass-metallicity relation which is an important outcome of feedback. The formation of such low-mass, but metal-rich and compact, objects is a long-standing puzzle. Using a pair of high-resolution N-body+gas simulations, we investigate the evolution of a gas-rich low-mass galaxy on a highly radial orbit around a massive host galaxy. As the infalling low-mass galaxy passes through the host's corona at supersonic speeds, its diffuse gas outskirts are stripped by ram pressure, as expected. However, the compactness increases rapidly because of bursty star formation in the gas tidally driven to the centre. The metal-rich gas produced by supernovae and stellar winds is confined by the ram pressure from the surrounding environment, leading to subsequent generations of stars being more metal-rich. After the gas is depleted, tidal interactions enhance the metallicity further via the stripping of weakly bound, old, and metal-poor stars, while the size of the satellite is changed only modestly. The outcome is a metal-rich cE that is consistent with observations. These results argue that classical cEs are neither the stripped remnants of much more massive galaxies nor the merger remnants of normal dwarfs. We present observable predictions that can be used to test our model.

astro-ph.GA↗

arXiv2019

The imprint of clump formation at high redshift. I. A disc alpha-abundance dichotomy

The disc structure of the Milky Way is marked by a chemical dichotomy, with high-alpha and low-alpha abundance sequences, traditionally identified with the geometric thick and thin discs. This identification is aided by the old ages of the high-alpha stars, and lower average ages of the low-alpha ones. Recent large scale surveys such as APOGEE have provided a wealth of data on this chemical structure, including showing that an identification of chemical and geometric thick discs is not exact, but the origin of the chemical dichotomy has remained unclear. Here we demonstrate that a dichotomy arises naturally if the early gas-rich disc fragments, leading to some fraction of the star formation occurring in clumps of the type observed in high-redshift galaxies. These clumps have high star formation rate density. They, therefore, enrich rapidly, moving from the low-alpha to the high-alpha sequence, while more distributed star formation produces the low-alpha sequence. We demonstrate that this model produces a chemically-defined thick disc that has many of the properties of the Milky Way's thick disc. Because clump formation is common in high redshift galaxies, we predict that chemical bimodalities are common in massive galaxies.

astro-ph.GA↗

arXiv2018

The Prevalence of Type III Disc Breaks in HI-rich and Low-spin Galaxies

We investigate the origin of the type III (up-bending) discs based on a sample of 1808 galaxies from SDSS and a sub-sample of 286 galaxies with HI data from ALFALFA. We examine how the type III fraction f3, the fraction of disc galaxies which host up-bending disc breaks, depends on other galactic properties. We confirm that f3 strongly depends on the stellar concentration of galaxies. We find that HI-rich galaxies with low spins tend to have significantly more type III disc breaks than control galaxies, which are matched in concentration and stellar mass. This trend is independent of the existence of strong bars or environment of the galaxies. This result is broadly consistent with predictions from theoretical simulations, and indicates in-situ star formation fueled by gas accretion to be an important mechanism that builds the outer discs of type III galaxies.

astro-ph.GA↗

arXiv2018

Curriculum Adversarial Training

Recently, deep learning has been applied to many security-sensitive applications, such as facial authentication. The existence of adversarial examples hinders such applications. The state-of-the-art result on defense shows that adversarial training can be applied to train a robust model on MNIST against adversarial examples; but it fails to achieve a high empirical worst-case accuracy on a more complex task, such as CIFAR-10 and SVHN. In our work, we propose curriculum adversarial training (CAT) to resolve this issue. The basic idea is to develop a curriculum of adversarial examples generated by attacks with a wide range of strengths. With two techniques to mitigate the forgetting and the generalization issues, we demonstrate that CAT can improve the prior art's empirical worst-case accuracy by a large margin of 25% on CIFAR-10 and 35% on SVHN. At the same, the model's performance on non-adversarial inputs is comparable to the state-of-the-art models.

cs.LG↗

arXiv2018

Shape of LOSVD\lowercase{s} in barred disks: Implications for future IFU surveys

The shape of LOSVDs (line-of-sight velocity distributions) carries important information about the internal dynamics of galaxies. The skewness of LOSVDs represents their asymmetric deviation from a Gaussian profile. Correlations between the skewness parameter ($h_3$) and the mean velocity (\vm) of a Gauss-Hermite series reflect the underlying stellar orbital configurations of different morphological components. Using two self-consistent $N$-body simulations of disk galaxies with different bar strengths, we investigate \hv\ correlations at different inclination angles. Similar to previous studies, we find anticorrelations in the disk area, and positive correlations in the bar area when viewed edge-on. However, at intermediate inclinations, the outer parts of bars exhibit anticorrelations, while the core areas dominated by the boxy/peanut-shaped (B/PS) bulges still maintain weak positive correlations. When viewed edge-on, particles in the foreground/background disk (the wing region) in the bar area constitute the main velocity peak, whereas the particles in the bar contribute to the high-velocity tail, generating the \hv\ correlation. If we remove the wing particles, the LOSVDs of the particles in the outer part of the bar only exhibit a low-velocity tail, resulting in a negative \hv\ correlation, whereas the core areas in the central region still show weakly positive correlations. We discuss implications for IFU observations on bars, and show that the variation of the \hv\ correlation in the disk galaxy may be used as a kinematic indicator of the bar and the B/PS bulge.

astro-ph.GA↗

arXiv2017

Black Hole Growth in Disk Galaxies Mediated by the Secular Evolution of Short Bars

The growth of black holes (BHs) in disk galaxies lacking classical bulges, which implies an absence of significant mergers, appears to be driven by secular processes. Short bars of sub-kiloparsec radius have been hypothesized to be an important mechanism for driving gas inflows to small scale, feeding central BHs. In order to quantify the maximum BH mass allowed by this mechanism, we examine the robustness of short bars to the dynamical influence of BHs. Large-scale bars are expected to be robust, long-lived structures; extremely massive BHs, which are rare, are needed to completely destroy such bars. However, we find that short bars, which are generally embedded in large-scale outer bars, can be destroyed quickly when BHs of mass $M_{\rm bh}\sim0.05-0.2\%$ of the total stellar mass ($M_\star$) are present. In agreement with this prediction, all galaxies observed to host short bars have BHs with a mass fraction less than $0.2\%M_\star$. Thus the dissolution of short inner bars is possible, perhaps even frequent, in the universe. An important implication of this result is that inner bar-driven gas inflows may be terminated when BHs grow to $\sim0.1\%M_\star$. We predict that $0.2\%M_\star$ is the maximum mass of BHs allowed if they are fed predominately via inner bars. This value matches well the maximum ratio of BH-to-host galaxy stellar mass observed in galaxies with pseudo-bulges and most nearby AGN host galaxies. This hypothesis provides a novel explanation for the lower $M_{\rm bh}/M_\star$ in galaxies which have avoided significant mergers compared with galaxies with classical bulges.

astro-ph.GA↗