arXiv subjects

Ziang Yan

Publications and source records attributed to Ziang Yan.

At least 55 records · Page 3Linked to original sources

arXiv2025

KiDS-Legacy: Angular galaxy clustering from deep surveys with complex selection effects

Photometric galaxy surveys, despite their limited resolution along the line of sight, encode rich information about the large-scale structure (LSS) of the Universe thanks to the high number density and extensive depth of the data. However, the complicated selection effects in wide and deep surveys can potentially cause significant bias in the angular two-point correlation function (2PCF) measured from those surveys. In this paper, we measure the 2PCF from the newly published KiDS-Legacy sample. Given an $r$-band $5σ$ magnitude limit of $24.8$ and survey footprint of $1347$ deg$^2$, it achieves an excellent combination of sky coverage and depth for such a measurement. We find that complex selection effects, primarily induced by varying seeing, introduce over-estimation of the 2PCF by approximately an order of magnitude. To correct for such effects, we apply a machine learning-based method to recover an organised random (OR) that presents the same selection pattern as the galaxy sample. The basic idea is to find the selection-induced clustering of galaxies using a combination of self-organising maps (SOMs) and hierarchical clustering (HC). This unsupervised machine learning method is able to recover complicated selection effects without specifying their functional forms. We validate this SOM+HC method on mock deep galaxy samples with realistic systematics and selections derived from the KiDS-Legacy catalogue. Using mock data, we demonstrate that the OR delivers unbiased 2PCF cosmological parameter constraints, removing the $27σ$ offset in the galaxy bias parameter that is recovered when adopting uniform randoms. Blinded measurements on the real KiDS-Legacy data show that the corrected 2PCF is robust to the SOM+HC configuration near the optimal set-up suggested by the mock tests.

astro-ph.CO↗

arXiv2024

Impact of survey spatial variability on galaxy redshift distributions and the cosmological $3\times2$-point statistics for the Rubin Legacy Survey of Space and Time (LSST)

We investigate the impact of spatial survey non-uniformity on the galaxy redshift distributions for forthcoming data releases of the Rubin Observatory Legacy Survey of Space and Time (LSST). Specifically, we construct a mock photometry dataset degraded by the Rubin OpSim observing conditions, and estimate photometric redshifts of the sample using a template-fitting photo-$z$ estimator, BPZ, and a machine learning method, FlexZBoost. We select the Gold sample, defined as $i<25.3$ for 10 year LSST data, with an adjusted magnitude cut for each year and divide it into five tomographic redshift bins for the weak lensing lens and source samples. We quantify the change in the number of objects, mean redshift, and width of each tomographic bin as a function of the coadd $i$-band depth for 1-year (Y1), 3-year (Y3), and 5-year (Y5) data. In particular, Y3 and Y5 have large non-uniformity due to the rolling cadence of LSST, hence provide a worst-case scenario of the impact from non-uniformity. We find that these quantities typically increase with depth, and the variation can be $10-40\%$ at extreme depth values. Using Y3 as an example, we propagate the variable depth effect to the weak lensing $3\times2$pt analysis, and assess the impact on cosmological parameters via a Fisher forecast. We find that galaxy clustering is most susceptible to variable depth, and non-uniformity needs to be mitigated below $3\%$ to recover unbiased cosmological constraints. There is little impact on galaxy-shear and shear-shear power spectra, given the expected LSST Y3 noise.

astro-ph.CO↗

arXiv2024

6x2pt: Forecasting gains from joint weak lensing and galaxy clustering analyses with spectroscopic-photometric galaxy cross-correlations

We explore the enhanced self-calibration of photometric galaxy redshift distributions, $n(z)$, through the combination of up to six two-point functions. Our $\rm 3\times2pt$ configuration is comprised of photometric shear, spectroscopic galaxy clustering, and spectroscopic-photometric galaxy-galaxy lensing (GGL). We further include spectroscopic-photometric cross-clustering; photometric GGL; and photometric auto-clustering, using the photometric shear sample as density tracer. We perform simulated likelihood forecasts of the cosmological and nuisance parameter constraints for Stage-III- and Stage-IV-like surveys. For the Stage-III-like case, we employ realistic but perturbed redshift distributions, and distinguish between "coherent" shifting in one direction, versus more internal scattering and full-shape errors. For perfectly known $n(z)$, a $\rm 6\times2pt$ analysis gains $\sim40\%$ in Figure of Merit (FoM) in the $S_8\equivσ_8\sqrt{Ω_{\rm m}/0.3}$ and $Ω_{\rm m}$ plane relative to the $\rm 3\times2pt$ analysis. If untreated, coherent and incoherent redshift errors lead to inaccurate inferences of $S_8$ and $Ω_{\rm m}$, respectively. Employing bin-wise scalar shifts $δ{z}_i$ in the tomographic mean redshifts reduces cosmological parameter biases, with a $\rm 6x2pt$ analysis constraining the shift parameters with $2-4$ times the precision of a photometric $\rm 3^{ph}\times2pt$ analysis. For the Stage-IV-like survey, a $\rm 6\times2pt$ analysis doubles the FoM($σ_8{-}Ω_{\rm m}$) compared to any $\rm 3\times2pt$ or $\rm 3^{ph}\times2pt$ analysis, and is only $8\%$ less constraining than if the $n(z)$ were perfectly known. A Gaussian mixture model for the $n(z)$ reduces mean-redshift errors and preserves the $n(z)$ shape. It also yields the most accurate and precise cosmological constraints for any $N\rm\times2pt$ configuration given $n(z)$ biases.

astro-ph.CO↗

arXiv2024

InternVideo2: Scaling Foundation Models for Multimodal Video Understanding

We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts. Code and models are available at https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/.

cs.CV↗

arXiv2024

Probabilistic Forward Modeling of Galaxy Catalogs with Normalizing Flows

Evaluating the accuracy and calibration of the redshift posteriors produced by photometric redshift (photo-z) estimators is vital for enabling precision cosmology and extragalactic astrophysics with modern wide-field photometric surveys. Evaluating photo-z posteriors on a per-galaxy basis is difficult, however, as real galaxies have a true redshift but not a true redshift posterior. We introduce PZFlow, a Python package for the probabilistic forward modeling of galaxy catalogs with normalizing flows. For catalogs simulated with PZFlow, there is a natural notion of "true" redshift posteriors that can be used for photo-z validation. We use PZFlow to simulate a photometric galaxy catalog where each galaxy has a redshift, noisy photometry, shape information, and a true redshift posterior. We also demonstrate the use of an ensemble of normalizing flows for photo-z estimation. We discuss how PZFlow will be used to validate the photo-z estimation pipeline of the Dark Energy Science Collaboration (DESC), and the wider applicability of PZFlow for statistical modeling of any tabular data.

astro-ph.IM↗

arXiv2024

The star formation, dust, and abundance of galaxies with unWISE-CIB cross-correlations

The cosmic infrared background (CIB) is the accumulated infrared (IR) radiation mainly from interstellar dust heated up by early stars. In this work, we measure the cross-correlation between galaxies from the unWISE catalog and the CIB maps from the Planck satellite to simultaneously constrain the cosmic star formation rate (SFR), dust spectral energy distribution (SED), and the halo occupation distribution (HOD). The unWISE galaxy catalog is divided into three tomographic bins centered at $z\sim 0.6, 1.1, 1.5$, and the CIB maps are at 353, 545, and 857 GHz. We measure the cross-correlations between these galaxy samples and CIB maps and get a 194$σ$ signal within an angular scale $100<\ell<2000$, from which we constrain two CIB halo models from previous literature and one new model. The SFR, SED, and HOD model parameters are constrained consistently among the three models. Specifically, the dust temperature at $z=0$ is constrained $T_0={21.14}^{+1.02}_{-1.34}$ K, which is slightly lower than $T_0=24.4\pm1.9$ K measured by the Planck collaboration. The halo mass that gives the most efficient star formation is around $10^{{11.79}^{+0.73}_{-0.86}}M_{\odot}$. From the model parameters, combined with the SFR density at $z=0$ synthesized from multi-wavelength observations, we break the degeneracy between SED and SFR and recover the cosmic star formation history that is consistent with multi-wavelength surveys. We also constrain the graybody SED model in agreement with previous measurements from infrared flux stacking. From the HOD constraints, we derive an increasing trend of galaxy linear bias along redshifts that agrees with the results from cross- and auto-correlation with unWise galaxies. This study indicates the power of using CIB-galaxy cross-correlation to study star formation, dust, and abundance of galaxies across cosmic time.

astro-ph.CO↗

arXiv2023

On constraining Cosmology and the Halo Mass Function with Weak Gravitational Lensing

The discrepancy between the weak lensing (WL) and the {\it Planck} measurements of $S_8$ has been a subject of several studies. These studies tend to show that a suppression of the amplitude of the mass power spectrum $P(k)$ at high $k$ could resolve it. The WL signal at small-scale is sensitive to various effects, such as baryonic effects and intrinsic alignment. The accuracy of $P(k)$ depends on the modelling precision of these effects. A common approach for calculating $P(k)$ relies on a halo model. Amongst the various components necessary for the construction of $P(k)$, the halo mass function (HMF) is an important one. Traditionally, the HMF has been assumed to follow a fixed model. Recent literature shows that baryonic physics, amongst several other factors, could affect the HMF. In this study, we investigate the impact of allowing the HMF to vary. This provides a way of testing the validity of the halo model-HMF calibration using data. We find that the {\it Planck} cosmology is not compatible with the vanilla HMF for both the DES-y3 and the KiDS-1000 data. When the cosmology and the HMF parameters are allowed to vary, the {\it Planck} cosmology is no longer in tension. The modified HMF predicts a matter power spectrum with a $\sim 25\%$ power loss at $k\sim 1~{\rm h/Mpc}$, in agreement with the recent studies. We show that Stage IV surveys will be able to measure the HMF parameters with a few percent accuracy.

astro-ph.CO↗

arXiv2023

KiDS-1000: cross-correlation with Planck cosmic microwave background lensing and intrinsic alignment removal with self-calibration

Galaxy shear - cosmic microwave background (CMB) lensing convergence cross-correlations contain additional information on cosmology to auto-correlations. While being immune to certain systematic effects, they are affected by the galaxy intrinsic alignments (IA). This may be responsible for the reported low lensing amplitude of the galaxy shear $\times$ CMB convergence cross-correlations, compared to the standard Planck $Λ$CDM (cosmological constant and cold dark matter) cosmology prediction. In this work, we investigate how IA affects the Kilo-Degree Survey (KiDS) galaxy lensing shear - Planck CMB lensing convergence cross-correlation and compare it to previous treatments with or without IA taken into consideration. More specifically, we compare marginalization over IA parameters and the IA self-calibration (SC) method (with additional observables defined only from the source galaxies) and prove that SC can efficiently break the degeneracy between the CMB lensing amplitude $A_{\rm lens}$ and the IA amplitude $A_{\rm IA}$. We further investigate how different systematics affect the resulting $A_{\rm IA}$ and $A_{\rm lens}$, and validate our results with the MICE2 simulation. We find that by including the SC method to constrain IA, the information loss due to the degeneracy between CMB lensing and IA is strongly reduced. The best-fit values are $A_{\rm lens}=0.84^{+0.22}_{-0.22}$ and $A_{\rm IA}=0.60^{+1.03}_{-1.03}$, while different angular scale cuts can affect $A_{\rm lens}$ by $\sim10\%$. We show that appropriate treatment of the boost factor, cosmic magnification, and photometric redshift modeling is important for obtaining the correct IA and cosmological results.

astro-ph.CO↗

arXiv2023

Exploring the mass and redshift dependence of the cluster pressure profile with stacks on thermal SZ maps

We provide novel constraints on the parameters defining the universal pressure profile (UPP) within clusters of galaxies, and explore their dependence on the cluster mass and redshift, from measurements of Sunyaev-Zel'dovich Compton-$y$ profiles. We employ both the $\textit{Planck}$ 2015 MILCA and the ACT-DR4 $y$ maps over the common $\sim 2,100\,\text{deg}^2$ footprint. We combine existing cluster catalogs based on KiDS, SDSS and DESI observations, for a total of 23,820 clusters spanning the mass range $10^{14.0}\,\text{M}_{\odot}<M_{500}<10^{15.1}\,\text{M}_{\odot}$ and the redshift range $0.02<z<0.98$. We split the clusters into three independent bins in mass and redshift; for each combination we detect the stacked SZ cluster signal and extract the mean $y$ angular profile. The latter is predicted theoretically adopting a halo model framework, and MCMCs are employed to estimate the UPP parameters, the hydrostatic mass bias $b_{\rm h}$ and possible cluster miscentering effects. We constrain $[P_0,c_{500},α,β]$ to $[5.9,2.0,1.8,4.9]$ with $\textit{Planck}$ and to $[3.8,1.3,1.0,4.4]$ with ACT using the full cluster sample, in agreement with previous findings. We do not find any compelling evidence for a residual mass or redshift dependence, thus expanding the validity of the cluster pressure profile over much larger $M_{500}$ and $z$ ranges; this is the first time the model has been tested on such a large (complete and representative) cluster sample. Finally, we obtain loose constraints on the hydrostatic mass bias in the range 0.2-0.3, again in broad agreement with previous works.

astro-ph.CO↗

arXiv2023

An End-to-End Framework for Marketing Effectiveness Optimization under Budget Constraint

Online platforms often incentivize consumers to improve user engagement and platform revenue. Since different consumers might respond differently to incentives, individual-level budget allocation is an essential task in marketing campaigns. Recent advances in this field often address the budget allocation problem using a two-stage paradigm: the first stage estimates the individual-level treatment effects using causal inference algorithms, and the second stage invokes integer programming techniques to find the optimal budget allocation solution. Since the objectives of these two stages might not be perfectly aligned, such a two-stage paradigm could hurt the overall marketing effectiveness. In this paper, we propose a novel end-to-end framework to directly optimize the business goal under budget constraints. Our core idea is to construct a regularizer to represent the marketing goal and optimize it efficiently using gradient estimation techniques. As such, the obtained models can learn to maximize the marketing goal directly and precisely. We extensively evaluate our proposed method in both offline and online experiments, and experimental results demonstrate that our method outperforms current state-of-the-art methods. Our proposed method is currently deployed to allocate marketing budgets for hundreds of millions of users on a short video platform and achieves significant business goal improvements. Our code will be publicly available.

cs.LG↗

arXiv2022

Cosmic star formation history with tomographic cosmic infrared background-galaxy cross-correlation

In this work, we probe the star formation history of the Universe using tomographic cross-correlation between the cosmic infrared background (CIB) and galaxy samples. The galaxy samples are from the Kilo-Degree Survey (KiDS), while the CIB maps are made from \planck\, sky maps. We measure the cross-correlation in harmonic space with a significance of 43$σ$. We model the cross-correlation with a halo model, which links CIB anisotropies to star formation rates (SFR) and galaxy abundance. We assume that SFR has a lognormal dependence on halo mass, while galaxy abundance follows the halo occupation distribution (HOD) model. The cross-correlations give a best-fit maximum star formation efficiency of $η_{\mathrm{max}}= 0.41^{+0.09}_{-0.14}$ at a halo mass $\log_{10}(M_{\mathrm{peak}}/M_{\odot})= {12.14\pm 0.36}$. The derived star formation rate density (SFRD) is well constrained up to $z\sim 1.5$. The constraining power at high redshift is mainly limited by the KiDS survey depth. A combination with external SFRD measurements from previous studies gives $\log_{10}(M_{\mathrm{peak}}/M_{\odot})=12.42^{+0.35}_{-0.19}$. This tightens the SFRD constraint up to $z=4$, yielding a peak SFRD of $0.09_{-0.004}^{+0.003}\,M_{\odot} \mathrm { year }^{-1} \mathrm{Mpc}^{-3}$ at $z=1.74^{+0.06}_{-0.02}$, corresponding to a lookback time of $10.05^{+0.12}_{-0.03}$ Gyr. Both constraints are consistent, and the derived SFRD agrees with previous studies and simulations. Additionally, we estimate the galaxy bias $b$ of KiDS galaxies from the constrained HOD parameters and yield an increasing bias from $b=1.1_{-0.31}^{+0.17}$ at $z=0$ to $b=1.96_{-0.64}^{+0.18}$ at $z=1.5$. Finally, we provide a forecast for future galaxy surveys and conclude that, due to their considerable depth, future surveys will yield a much tighter constraint on the evolution of the SFRD.

astro-ph.GA↗

arXiv2022

Sill-Net: Feature Augmentation with Separated Illumination Representation

For visual object recognition tasks, the illumination variations can cause distinct changes in object appearance and thus confuse the deep neural network based recognition models. Especially for some rare illumination conditions, collecting sufficient training samples could be time-consuming and expensive. To solve this problem, in this paper we propose a novel neural network architecture called Separating-Illumination Network (Sill-Net). Sill-Net learns to separate illumination features from images, and then during training we augment training samples with these separated illumination features in the feature space. Experimental results demonstrate that our approach outperforms current state-of-the-art methods in several object classification benchmarks.

cs.CV↗

arXiv2022

Dynamical cluster masses from photometric surveys

The masses of galaxy clusters can be measured using data obtained exclusively from wide photometric surveys in one of two ways: directly from the amplitude of the weak lensing signal or, indirectly, through the use of scaling relations calibrated using binned lensing measurements. In this paper, we build on a recently proposed idea and implement an alternative method based on the radial profile of the satellite distribution. This technique relies on splashback, a feature associated with the apocenter of recently accreted galaxies that offers a clear window into the phase-space structure of clusters without the use of velocity information. We carry out this dynamical measurement using the stacked satellite distribution around a sample of luminous red galaxies in the fourth data release of the Kilo-Degree Survey and validate our results using abundance-matching and lensing masses. To illustrate the power of this measurement, we combine dynamical and lensing mass estimates to robustly constrain scalar-tensor theories of gravity at cluster scales. Our results exclude departures from General Relativity of order unity. We conclude the paper by discussing the implications for future data sets. Because splashback mass measurements scale only with the survey volume, stage-IV photometric surveys are well-positioned to use splashback to provide high-redshift cluster masses.

astro-ph.CO↗

arXiv2022

Joint constraints on cosmology and the impact of baryon feedback: combining KiDS-1000 lensing with the thermal Sunyaev-Zeldovich effect from Planck and ACT

We conduct a pseudo-$C_\ell$ analysis of the tomographic cross-correlation between 1000 deg$^2$ of weak lensing data from the Kilo-Degree Survey (KiDS-1000) and the thermal Sunyaev-Zeldovich (tSZ) effect measured by Planck and the Atacama Cosmology Telescope (ACT). Using HMx, a halo-model-based approach that consistently models the gas, star, and dark matter components, we are able to derive constraints on both cosmology and baryon feedback for the first time from these data, marginalising over redshift uncertainties, intrinsic alignment of galaxies, and contamination by the cosmic infrared background (CIB). We find our results to be insensitive to the CIB, while intrinsic alignment provides a small but significant contribution to the lensing--tSZ cross-correlation. The cosmological constraints are consistent with those of other low-redshift probes and prefer strong baryon feedback. The inferred amplitude of the lensing--tSZ cross-correlation signal, which scales as $σ_8(Ω_\mathrm{m}/0.3)^{0.2}$, is low by $\sim 2\,σ$ compared to the primary cosmic microwave background constraints by Planck. The lensing--tSZ measurements are then combined with pseudo-$C_\ell$ measurements of KiDS-1000 cosmic shear into a novel joint analysis, accounting for the full cross-covariance between the probes, providing tight cosmological constraints by breaking parameter degeneracies inherent to both probes. The joint analysis gives an improvement of 40% on the constraint of $S_8=σ_8\sqrt{Ω_\mathrm{m}/0.3}$ over cosmic shear alone, while providing constraints on baryon feedback consistent with hydrodynamical simulations, demonstrating the potential of such joint analyses with baryonic tracers such as the tSZ effect. We discuss remaining modelling challenges that need to be addressed if these baryonic probes are to be included in future precision-cosmology analyses.

astro-ph.CO↗

arXiv2021

Probing galaxy bias and intergalactic gas pressure with KiDS Galaxies-tSZ-CMB lensing cross-correlations

We constrain the redshift dependence of gas pressure bias $\left\langle b_{y} P_{\mathrm{e}}\right\rangle$ (bias-weighted average electron pressure), which characterises the thermodynamics of intergalactic gas, through a combination of cross-correlations between galaxy positions and the thermal Sunyaev-Zeldovich (tSZ) effect, as well as galaxy positions and the gravitational lensing of the cosmic microwave background (CMB). The galaxy sample is from the fourth data release of the Kilo-Degree Survey (KiDS). The tSZ $y$ map and the CMB lensing map are from the {\textit{Planck}} 2015 and 2018 data releases, respectively. The measurements are performed in five redshift bins with $z\lesssim1$. With these measurements, combining galaxy-tSZ and galaxy-CMB lensing cross-correlations allows us to break the degeneracy between galaxy bias and gas pressure bias, and hence constrain them simultaneously. In all redshift bins, the best-fit values of $\bpe$ are at a level of $\sim 0.3\, \mathrm{meV/cm^3}$ and increase slightly with redshift. The galaxy bias is consistent with unity in all the redshift bins. Our results are not sensitive to the non-linear details of the cross-correlation, which are smoothed out by the {\textit{Planck}} beam. Our measurements are in agreement with previous measurements as well as with theoretical predictions. We also show that our conclusions are not changed when CMB lensing is replaced by galaxy lensing, which shows the consistency of the two lensing signals despite their radically different redshift ranges. This study demonstrates the feasibility of using CMB lensing to calibrate the galaxy distribution such that the galaxy distribution can be used as a mass proxy without relying on the precise knowledge of the matter distribution.

astro-ph.CO↗

arXiv2020

Learning Fast Approximations of Sparse Nonlinear Regression

The idea of unfolding iterative algorithms as deep neural networks has been widely applied in solving sparse coding problems, providing both solid theoretical analysis in convergence rate and superior empirical performance. However, for sparse nonlinear regression problems, a similar idea is rarely exploited due to the complexity of nonlinearity. In this work, we bridge this gap by introducing the Nonlinear Learned Iterative Shrinkage Thresholding Algorithm (NLISTA), which can attain a linear convergence under suitable conditions. Experiments on synthetic data corroborate our theoretical results and show our method outperforms state-of-the-art methods.

cs.LG↗

arXiv2020

FMA-ETA: Estimating Travel Time Entirely Based on FFN With Attention

Estimated time of arrival (ETA) is one of the most important services in intelligent transportation systems and becomes a challenging spatial-temporal (ST) data mining task in recent years. Nowadays, deep learning based methods, specifically recurrent neural networks (RNN) based ones are adapted to model the ST patterns from massive data for ETA and become the state-of-the-art. However, RNN is suffering from slow training and inference speed, as its structure is unfriendly to parallel computing. To solve this problem, we propose a novel, brief and effective framework mainly based on feed-forward network (FFN) for ETA, FFN with Multi-factor self-Attention (FMA-ETA). The novel Multi-factor self-attention mechanism is proposed to deal with different category features and aggregate the information purposefully. Extensive experimental results on the real-world vehicle travel dataset show FMA-ETA is competitive with state-of-the-art methods in terms of the prediction accuracy with significantly better inference speed.

cs.LG↗

arXiv2019

Adversarial Margin Maximization Networks

The tremendous recent success of deep neural networks (DNNs) has sparked a surge of interest in understanding their predictive ability. Unlike the human visual system which is able to generalize robustly and learn with little supervision, DNNs normally require a massive amount of data to learn new concepts. In addition, research works also show that DNNs are vulnerable to adversarial examples-maliciously generated images which seem perceptually similar to the natural ones but are actually formed to fool learning models, which means the models have problem generalizing to unseen data with certain type of distortions. In this paper, we analyze the generalization ability of DNNs comprehensively and attempt to improve it from a geometric point of view. We propose adversarial margin maximization (AMM), a learning-based regularization which exploits an adversarial perturbation as a proxy. It encourages a large margin in the input space, just like the support vector machines. With a differentiable formulation of the perturbation, we train the regularized DNNs simply through back-propagation in an end-to-end manner. Experimental results on various datasets (including MNIST, CIFAR-10/100, SVHN and ImageNet) and different DNN architectures demonstrate the superiority of our method over previous state-of-the-arts. Code and models for reproducing our results will be made publicly available.

cs.LG↗