arXiv subjects

Yu Liu

Publications and source records attributed to Yu Liu.

At least 1,009 records · Page 56Linked to original sources

arXiv2019

Knowledge Distillation via Route Constrained Optimization

Distillation-based learning boosts the performance of the miniaturized neural network based on the hypothesis that the representation of a teacher model can be used as structured and relatively weak supervision, and thus would be easily learned by a miniaturized model. However, we find that the representation of a converged heavy model is still a strong constraint for training a small student model, which leads to a high lower bound of congruence loss. In this work, inspired by curriculum learning we consider the knowledge distillation from the perspective of curriculum learning by routing. Instead of supervising the student model with a converged teacher model, we supervised it with some anchor points selected from the route in parameter space that the teacher model passed by, as we called route constrained optimization (RCO). We experimentally demonstrate this simple operation greatly reduces the lower bound of congruence loss for knowledge distillation, hint and mimicking learning. On close-set classification tasks like CIFAR100 and ImageNet, RCO improves knowledge distillation by 2.14% and 1.5% respectively. For the sake of evaluating the generalization, we also test RCO on the open-set face recognition task MegaFace.

cs.LG↗

arXiv2019

Measuring Road Network Topology Vulnerability by Ricci Curvature

Describing the basic properties of road network systems, such as their robustness, vulnerability, and reliability, has been a very important research topic in the field of urban transportation. Current research mainly uses several statistical indicators of complex networks to analyze the road network systems. However, these methods are essentially node-based. These node-based methods are more concerned with the number of connections between nodes, and lack of consideration for interactions. So, this leads to the well-known node paradox problem, and their ability of characterizing the local and intrinsic properties of a network is weak. From the perspective of network intrinsic geometry, this paper proposes a method for measuring road network vulnerability using a discrete Ricci curvature, which can identify the key sections of a road network and indicate its fragile elements. The results show that our method performs better than complex network statistics on measuring the vulnerability of a road network. Additionally, it can characterize the evolution of the road network vulnerability among different periods of time in the same city through our method. Finally, we compare our method with the previous method of centrality and show the different between them. This article provides a new perspective on a geometry to analyze the vulnerability of a road network and describes the inherent nature of the vulnerability of a road system from a new perspective. It also contributes to enriching the analytical methods of complex road networks.

cs.SI↗

arXiv2019

Conditional Adversarial Generative Flow for Controllable Image Synthesis

Flow-based generative models show great potential in image synthesis due to its reversible pipeline and exact log-likelihood target, yet it suffers from weak ability for conditional image synthesis, especially for multi-label or unaware conditions. This is because the potential distribution of image conditions is hard to measure precisely from its latent variable $z$. In this paper, based on modeling a joint probabilistic density of an image and its conditions, we propose a novel flow-based generative model named conditional adversarial generative flow (CAGlow). Instead of disentangling attributes from latent space, we blaze a new trail for learning an encoder to estimate the mapping from condition space to latent space in an adversarial manner. Given a specific condition $c$, CAGlow can encode it to a sampled $z$, and then enable robust conditional image synthesis in complex situations like combining person identity with multiple attributes. The proposed CAGlow can be implemented in both supervised and unsupervised manners, thus can synthesize images with conditional information like categories, attributes, and even some unknown properties. Extensive experiments show that CAGlow ensures the independence of different conditions and outperforms regular Glow to a significant extent.

cs.CV↗

arXiv2019

Correlation Congruence for Knowledge Distillation

Most teacher-student frameworks based on knowledge distillation (KD) depend on a strong congruent constraint on instance level. However, they usually ignore the correlation between multiple instances, which is also valuable for knowledge transfer. In this work, we propose a new framework named correlation congruence for knowledge distillation (CCKD), which transfers not only the instance-level information, but also the correlation between instances. Furthermore, a generalized kernel method based on Taylor series expansion is proposed to better capture the correlation between instances. Empirical experiments and ablation studies on image classification tasks (including CIFAR-100, ImageNet-1K) and metric learning tasks (including ReID and Face Recognition) show that the proposed CCKD substantially outperforms the original KD and achieves state-of-the-art accuracy compared with other SOTA KD-based methods. The CCKD can be easily deployed in the majority of the teacher-student framework such as KD and hint-based learning methods.

cs.CV↗

arXiv2019

Pulsed Quantum-State Reconstruction of Dark Systems

We propose a novel strategy to reconstruct the quantum state of dark systems, i.e., degrees of freedom that are not directly accessible for measurement or control. Our scheme relies on the quantum control of a two-level probe that exerts a state-dependent potential on the dark system. Using a sequence of control pulses applied to the probe makes it possible to tailor the information one can obtain and, for example, allows us to reconstruct the density operator of a dark spin as well as the Wigner characteristic function of a harmonic oscillator. Because of the symmetry of the applied pulse sequence, this scheme is robust against slow noise on the probe. The proof-of-principle experiments are readily feasible in solid-state spins and trapped ions.

quant-ph↗

arXiv2019

Possible correlations between gamma-ray burst and its host galaxy offset

We collected the information of 304 gamma-ray bursts (GRBs) from the literature, and analyzed the correlations among the host galaxy offsets (the distance from the site of the GRB to the center of its host galaxy), $T_{\rm 90,i}$ (the duration $T_{\rm 90}$ in rest-frame), $T_{\rm R45,i}$ (the duration $T_{\rm R45}$ in rest-frame), $E_{\rm γ,iso}$ (the isotropic equivalent energy), $L_{\rm γ,iso}$ ($=E_{\rm γ,iso}/T_{\rm 90,i}$, the isotropic equivalent luminosity) and $L_{\rm pk}$ (peak luminosity). We found that $ T_{\rm 90,i}$, $ T_{\rm R45,i}$, $ E_{\rm γ,iso}$, $L_{\rm pk}$ have negative correlation with $ {\rm offset}$, which is consistent with origin of short GRBs (SGRBs) and long GRBs (LGRBs). On separate analysis, we found similar results for $\log E_{\rm γ,iso}$ - $\log {\rm (offset)}$ and $\log L_{\rm pk}$ - $\log {\rm (offset)}$ relations in case of SGRBs only, while no obvious relation for LGRBs. There is no correlations between offset and $L_{\rm γ,iso}$. We also put the special GRB 170817A {and GRB 060218A} on the plots. {The two GRBs both have low luminosity and small offset.} In the $ \log ({\rm offset})- \log T_{\rm 90,i}$ plot, we found GRB 170817A locates in between the two regions of SGRBs and LGRBs and it is the outlier in the $ {\rm offset}-E_{\rm γ, iso}$, $ {\rm offset}-L_{\rm γ, iso}$ and $ {\rm offset}-L_{\rm pk}$ plots. Together with GRB 060218A being an outlier in all plots, it indicates the speciality of GRBs 170817A and 060218A, and might imply more subgroups of the GRB samples.

astro-ph.HE↗

arXiv2019

Abelian quotients associated with fully rigid subcategories

In this article, we study the Gorenstein property of abelian quotient categories induced by fully rigid subcategories on an exact category B. We also study when d-cluster tilting subcategories become fully rigid. We show that the quotient abelian category induced by such d-cluster tilting subcategories are hereditary.

math.RT↗

arXiv2019

A comprehensive statistical study on gamma-ray bursts

In order to obtain an overview of the gamma-ray bursts (GRBs), we need a full sample. In this paper, we collected 6289 GRBs (from GRB 910421 to GRB 160509A) from the literature, including prompt emission, afterglow and host galaxy properties. We hope to use this large sample to reveal the intrinsic properties of GRB. We have listed all the data in machine readable tables, including the properties of the GRBs, correlation coefficients and linear regression results of two arbitrary parameters, and linear regression results of any three parameters. These machine readable tables could be used as a data reservoir for further studies on the classifications or correlations. One may find some intrinsic properties from these statistical results. With this comprehensive table, it is possible to find relations between different parameters, and to classify the GRBs into different kinds of sub-groups. With the completion, it may reveal the nature of GRBs and may be used as tools like pseudo-redshift indicators, standard candles, etc. All the machine readable data and statistical results are available on the website of the journal.

astro-ph.HE↗

arXiv2019

First Unambiguous Imaging of Large-Scale Quasi-Periodic Extreme-Ultraviolet Wave or Shock

We report the first unambiguous quasi-periodic large-scale extreme-ultraviolet (EUV) wave or shock that was detected by the Atmospheric Imaging Assembly on board the Solar Dynamics Observatory. During the whip-like unwinding eruption of a small filament on 2012 April 24, multiple consecutive large-scale wavefronts emanating from AR11467 were observed simultaneously along the solar surface and a closed transequatorial loop system. In the meantime, an upward propagating dome-shaped wavefront was also observed, whose initial speed and deceleration are about 1392 km/s and 1.78 km/s^2, respectively. Along the solar surface, the quasi-peridoic wavefronts had a period of about 163 +/- 21 seconds and propagated at a nearly constant speed of 747 +/- 26 km/s; they interacted with active region AR11469 and launched a sympathetic upward propagating secondary EUV wave. The wavefronts along the loop system propagated at a speed of 897 km/s, and they were reflected back at the southern end of the loop system at a similar speed. In addition to the propagating waves, a standing kink wave was also present in the loop system simultaneously. Periodicity analysis reveals that the period of the wavefronts was consistent with that of the unwinding helical structures of the erupting filament. Based on these observational facts, we propose that the observed quasi-periodic EUV wavefronts were most likely excited by the periodic unwinding motion of the filament helical structures. In addition, two different seismological methods are applied to derive the magnetic field strength of the loop system, and for the first time the reliability of these inversion techniques are tested with the same magnetic structure.

astro-ph.SR↗

arXiv2019

Learning Pairwise Relationship for Multi-object Detection in Crowded Scenes

As the post-processing step for object detection, non-maximum suppression (GreedyNMS) is widely used in most of the detectors for many years. It is efficient and accurate for sparse scenes, but suffers an inevitable trade-off between precision and recall in crowded scenes. To overcome this drawback, we propose a Pairwise-NMS to cure GreedyNMS. Specifically, a pairwise-relationship network that is based on deep learning is learned to predict if two overlapping proposal boxes contain two objects or zero/one object, which can handle multiple overlapping objects effectively. Through neatly coupling with GreedyNMS without losing efficiency, consistent improvements have been achieved in heavily occluded datasets including MOT15, TUD-Crossing and PETS. In addition, Pairwise-NMS can be integrated into any learning based detectors (Both of Faster-RCNN and DPM detectors are tested in this paper), thus building a bridge between GreedyNMS and end-to-end learning detectors.

cs.CV↗

arXiv2019

Anisotropic magnetic entropy change in Cr$_2$X$_2$Te$_6$ (X = Si and Ge)

Intrinsic, two-dimensional (2D) ferromagnetic semiconductors are an important class of materials for spintronics applications. Cr$_2$X$_2$Te$_6$ (X = Si and Ge) semiconductors show 2D Ising-like ferromagnetism, which is preserved in few-layer devices. The maximum magnetic entropy change associated with the critical properties around the ferromagnetic transition for Cr$_2$Si$_2$Te$_6$ $-ΔS_M^{max} \sim$ 5.05 J kg$^{-1}$ K$^{-1}$ is much larger than $-ΔS_M^{max} \sim$ 2.64 J kg$^{-1}$ K$^{-1}$ for Cr$_2$Ge$_2$Te$_6$ with an out-of-plane field change of 5 T. The rescaled $-ΔS_M(T,H)$ curves collapse onto a universal curve independent of temperature and field for both materials. This indicates similar critical behavior and 2D Ising magnetism, confirming the magnetocrystalline anisotropy that could preserve the long-range ferromagnetism in few-layers of Cr$_2$X$_2$Te$_6$.

cond-mat.str-el↗

arXiv2018

T-GCN: A Temporal Graph ConvolutionalNetwork for Traffic Prediction

Accurate and real-time traffic forecasting plays an important role in the Intelligent Traffic System and is of great significance for urban traffic planning, traffic management, and traffic control. However, traffic forecasting has always been considered an open scientific issue, owing to the constraints of urban road network topological structure and the law of dynamic change with time, namely, spatial dependence and temporal dependence. To capture the spatial and temporal dependence simultaneously, we propose a novel neural network-based traffic forecasting method, the temporal graph convolutional network (T-GCN) model, which is in combination with the graph convolutional network (GCN) and gated recurrent unit (GRU). Specifically, the GCN is used to learn complex topological structures to capture spatial dependence and the gated recurrent unit is used to learn dynamic changes of traffic data to capture temporal dependence. Then, the T-GCN model is employed to traffic forecasting based on the urban road network. Experiments demonstrate that our T-GCN model can obtain the spatio-temporal correlation from traffic data and the predictions outperform state-of-art baselines on real-world traffic datasets. Our tensorflow implementation of the T-GCN is available at https://github.com/lehaifeng/T-GCN.

cs.LG↗

arXiv2018

Joint Beam and Channel Tracking for Two-Dimensional Phased Antenna Arrays

Analog beamforming is a low-cost architecture for millimeter-wave (mmWave) mobile communications. However, it has two disadvantages for serving fast mobility users: (i) the mmWave beam in the wireless channel and the beam steered by analog beamforming have small angular spreads which are difficult to align with each other and (ii) the receiver can only observe the mmWave channel in one beam direction and rely on beam-probing algorithms to check other directions. In this paper, we develop a beam probing and tracking algorithm that can efficiently track fast-moving mmWave beams in three-dimensional (3D) space. This algorithm has several salient features: (1) fading channel supportive: it can simultaneously track the channel coefficient and two-dimensional (2D) beam direction in fading channel environments; (2) low probing overhead: it achieves the minimum probing requirement for joint beam and channel tracking; (3) fast tracking speed and high tracking accuracy: its tracking error converges to the minimum Cramer-Rao lower bound (CRLB) in static scenarios in theory and it outperforms several existing tracking algorithms with lower tracking error and faster tracking speed in simulations.

eess.SP↗

arXiv2018

$n$-exangulated categories

For each positive integer $n$ we introduce the notion of $n$-exangulated categories as higher dimensional analogues of extriangulated categories defined by Nakaoka-Palu. We characterize which $n$-exangulated categories are $n$-exact in the sense of Jasso and which are $(n+2)$-angulated in the sense of Geiss-Keller-Oppermann. For extriangulated categories with enough projectives and injectives we introduce the notion of $n$-cluster tilting subcategories and show that under certain conditions such $n$-cluster tilting subcategories are $n$-exangulated.

math.CT↗

arXiv2018

RGB-D Based Action Recognition with Light-weight 3D Convolutional Networks

Different from RGB videos, depth data in RGB-D videos provide key complementary information for tristimulus visual data which potentially could achieve accuracy improvement for action recognition. However, most of the existing action recognition models solely using RGB videos limit the performance capacity. Additionally, the state-of-the-art action recognition models, namely 3D convolutional neural networks (3D-CNNs) contain tremendous parameters suffering from computational inefficiency. In this paper, we propose a series of 3D light-weight architectures for action recognition based on RGB-D data. Compared with conventional 3D-CNN models, the proposed light-weight 3D-CNNs have considerably less parameters involving lower computation cost, while it results in favorable recognition performance. Experimental results on two public benchmark datasets show that our models can approximate or outperform the state-of-the-art approaches. Specifically, on the RGB+D-NTU (NTU) dataset, we achieve 93.2% and 97.6% for cross-subject and cross-view measurement, and on the Northwestern-UCLA Multiview Action 3D (N-UCLA) dataset, we achieve 95.5% accuracy of cross-view.

cs.CV↗

arXiv2018

How far from automatically interpreting deep learning

In recent years, deep learning researchers have focused on how to find the interpretability behind deep learning models. However, today cognitive competence of human has not completely covered the deep learning model. In other words, there is a gap between the deep learning model and the cognitive mode. How to evaluate and shrink the cognitive gap is a very important issue. In this paper, the interpretability evaluation, the relationship between the generalization performance and the interpretability of the model and the method for improving the interpretability are concerned. A universal learning framework is put forward to solve the equilibrium problem between the two performances. The uniqueness of solution of the problem is proved and condition of unique solution is obtained. Probability upper bound of the sum of the two performances is analyzed.

cs.LG↗

arXiv2018

Anomalous Hall effect in trigonal Cr$_5$Te$_8$ single crystal

We report anomalous Hall effect (AHE) and transport properties of trigonal Cr$_5$Te$_8$ (tr-Cr$_5$Te$_8$) single crystals. The electrical resistivity as well as the Seebeck coefficient shows a clear kink at the paramagnetic-ferromagnetic transition of tr-Cr$_5$Te$_8$, which is also confirmed by the heat capacity measurement. The scaling behavior between anomalous Hall resistivity $ρ^A_{xy}$ and longitudinal resistivity $ρ_{xx}$ is linear below $T_c$. Further analysis suggests that the AHE in tr-Cr$_5$Te$_8$ is dominated by the skew-scattering mechanism rather than the intrinsic or extrinsic side-jump mechanism.

cond-mat.str-el↗

arXiv2018

Gradient Harmonized Single-stage Detector

Despite the great success of two-stage detectors, single-stage detector is still a more elegant and efficient way, yet suffers from the two well-known disharmonies during training, i.e. the huge difference in quantity between positive and negative examples as well as between easy and hard examples. In this work, we first point out that the essential effect of the two disharmonies can be summarized in term of the gradient. Further, we propose a novel gradient harmonizing mechanism (GHM) to be a hedging for the disharmonies. The philosophy behind GHM can be easily embedded into both classification loss function like cross-entropy (CE) and regression loss function like smooth-$L_1$ ($SL_1$) loss. To this end, two novel loss functions called GHM-C and GHM-R are designed to balancing the gradient flow for anchor classification and bounding box refinement, respectively. Ablation study on MS COCO demonstrates that without laborious hyper-parameter tuning, both GHM-C and GHM-R can bring substantial improvement for single-stage detector. Without any whistles and bells, our model achieves 41.6 mAP on COCO test-dev set which surpasses the state-of-the-art method, Focal Loss (FL) + $SL_1$, by 0.8.

cs.CV↗