arXiv subjects

Yao Chen

Publications and source records attributed to Yao Chen.

At least 163 records · Page 9Linked to original sources

arXiv2020

Effective Algorithm-Accelerator Co-design for AI Solutions on Edge Devices

High quality AI solutions require joint optimization of AI algorithms, such as deep neural networks (DNNs), and their hardware accelerators. To improve the overall solution quality as well as to boost the design productivity, efficient algorithm and accelerator co-design methodologies are indispensable. In this paper, we first discuss the motivations and challenges for the Algorithm/Accelerator co-design problem and then provide several effective solutions. Especially, we highlight three leading works of effective co-design methodologies: 1) the first simultaneous DNN/FPGA co-design method; 2) a bi-directional lightweight DNN and accelerator co-design method; 3) a differentiable and efficient DNN and accelerator co-search method. We demonstrate the effectiveness of the proposed co-design approaches using extensive experiments on both FPGAs and GPUs, with comparisons to existing works. This paper emphasizes the importance and efficacy of algorithm-accelerator co-design and calls for more research breakthroughs in this interesting and demanding area.

cs.AR↗

arXiv2020

Microwave Study of a Solar Circular Ribbon Flare

A circular ribbon flare SOL2014-12-17T04:51 is studied using the 17/34 GHz maps from the Nobeyama Radioheliograph (NoRH) along with (E)UV and magnetic data from the Solar Dynamics Observatory (SDO). We report the following three findings as important features of the microwave CRF. (1) The first preflare activation comes in the form of a gradual increase of the 17 GHz flux without a counterpart at 34 GHz, which indicates thermal preheating. The first sign of nonthermal activity occurs in the form of stepwise flux increases at both 17 and 34 GHz about 4 min before the impulsive phase. (2) Until the impulsive phase, the microwave emission over the entire active region is in a single polarization state matching the magnetic polarity of the surrounding fields. During and after the impulsive phase, the sign of the 17 GHz polarization state reverses in the core region, which implies a magnetic breakout--type eruption in a fan-spine magnetic structure. (3) The 17 GHz flux around the time of the eruption shows quasi-periodic variations with periods of 1--2 min. The pre-eruption oscillation is more obvious in total intensity at one end of the flare loop, and the post-eruption oscillation, more obvious in the polarized intensity at a region near the inner spine. We interpret this transition as the transfer of oscillatory power from kink mode oscillation to torsional Alfvén waves propagating along the spine field after the eruption. We argue that these three processes are inter-related and indicate a breakout process in a fan-spine structure.

astro-ph.SR↗

arXiv2020

Formation of a tiny flux rope in the center of an active region driven by magnetic flux emergence, convergence, and cancellation

Flux ropes are generally believed to be core structures of solar eruptions that are significant for the space weather, but their formation mechanism remains intensely debated. We report on the formation of a tiny flux rope beneath clusters of active region loops on 2018 August 24. Combining the high-quality multiwavelength observations from multiple instruments, we studied the event in detail in the photosphere, chromosphere, and corona. In the source region, the continual emergence of two positive polarities (P1 and P2) that appeared as two pores (A and B)is unambiguous. Interestingly, P2 and Pore B slowly approached P1 and Pore A, implying a magnetic flux convergence. During the emergence and convergence, P1 and P2 successively interacted with a minor negative polarity (N3) that emerged, which led to a continuous magnetic flux cancellation. As a result, the overlying loops became much sheared and finally evolved into a tiny twisted flux rope that was evidenced by a transient inverse S-shaped sigmoid, the twisted filament threads with blueshift and redshift signatures, and a hot channel. All the results show that the formation of the tiny flux rope in the center of the active region was closely associated with the continuous magnetic flux emergence, convergence, and cancellation in the photosphere. Hence, we suggest that the magnetic flux emergence, convergence, and cancellation are crucial for the formation of the tiny flux rope.

astro-ph.SR↗

arXiv2020

Do All Interplanetary Coronal Mass Ejections Have A Magnetic Flux Rope Structure Near 1 AU?

Interplanetary coronal mass ejections (ICMEs) often consist of a shock wave, sheath region, and ejecta region. The ejecta regions are divided into two broad classes: magnetic clouds (MC) that exhibit the characteristics of magnetic flux ropes and non-magnetic clouds (NMC) that do not. As CMEs result from eruption of magnetic flux ropes, it is important to answer why NMCs do not have the flux rope features. One claims that NMCs lose their original flux rope features due to the interactions between ICMEs or ICMEs and other large scale structures during their transit in the heliosphere. The other attributes this phenomenon to the geometric selection effect, i.e., when an ICME has its nose (flank, including leg and non-leg flanks) pass through the observing spacecraft, the MC (NMC) features will be detected along the spacecraft trajectory within the ejecta. In this Letter, we examine which explanation is more reasonable through the geometric properties of ICMEs. If the selection effect leads to different ejecta types, MCs should have narrower sheath region compared to NMCs from the statistical point of view, which is confirmed by our statistics. Besides, we find that NMCs have the similar size in solar cycles 23 and 24, and NMCs are smaller than MCs in cycle 23 but larger than MCs in cycle 24. This suggests that most NMCs have their leg flank pass through the spacecraft. Our geometric analyses support that all ICMEs should have a magnetic flux rope structure near 1 AU.

astro-ph.SR↗

arXiv2020

Phonon-induced anomalous gauge potential for photonic isolation in frequency space

Photonic gauge potentials are crucial for manipulating charge-neutral photons like their counterpart electrons in the electromagnetic field, allowing analogous Aharonov-Bohm effect in photonics and paving the way for critical applications like photonic isolation. Normally, a gauge potential exhibits phase inversion along two opposite propagation paths. Here we experimentally demonstrate phonon-induced anomalous gauge potentials with non-inverted gauge phases in a spatial-frequency space, where quasi-phase-matched nonlinear Brillouin scatterings enable such unique direction-dependent gauge phases. Based on this scheme, we construct photonic isolators in the frequency domain permitting nonreciprocal propagation of light along the frequency axis, where coherent phase control in the photonic isolator allows switching completely the directionality through an Aharonov-Bohm interferometer. Moreover, similar coherent controlled unidirectional frequency conversions are also illustrated. These results may offer a unique platform for a compact, integrated solution to implement synthetic-dimension devices for on-chip optical signal processing.

physics.optics↗

arXiv2020

The initiation of a solar streamer blowout coronal mass ejection arising from the streamer flank

Streamer blowout (SBO) coronal mass ejections (CMEs) represent a particular class of CMEs that are characterized by a gradual swelling of the overlying streamer and a slow CME containing a flux-rope structure. SBO CMEs arising from the streamer flank fall into a special category of SBO CMEs involving three lower arches under the higher streamer arcade. However, the initiation mechanism for this special category of SBO CMEs remains elusive, due to the observational limitations. Here we report critical observations of a SBO CME associated with the eruption of a polar crown filament that originated from the streamer flank. The filament slowly rose toward the solar equator with the writhing motion, and underwent a sudden acceleration before its eruption. Interestingly, during the rising, the filament fields experienced gradual external reconnections, which is evidenced by the dip-shaped bottom of the enveloping flux-rope structure changing from a smooth concave, the slow inflows ($\sim$1.8 km s$^{-1}$) from both the filament fields and the coronal loops beneath, and the persistent brightenings around the interface between the filament fields and the coronal loops beneath. The newly formed lower loops at the filament source and the Y-shaped structure in the stretched tail fields indicate the internal reconnections for the filament eruption. The clear signatures of the external and internal reconnections shed light on the initiation mechanisms of SBO CMEs.

astro-ph.SR↗

arXiv2020

VecQ: Minimal Loss DNN Model Compression With Vectorized Weight Quantization

Quantization has been proven to be an effective method for reducing the computing and/or storage cost of DNNs. However, the trade-off between the quantization bitwidth and final accuracy is complex and non-convex, which makes it difficult to be optimized directly. Minimizing direct quantization loss (DQL) of the coefficient data is an effective local optimization method, but previous works often neglect the accurate control of the DQL, resulting in a higher loss of the final DNN model accuracy. In this paper, we propose a novel metric called Vector Loss. Based on this new metric, we develop a new quantization solution called VecQ, which can guarantee minimal direct quantization loss and better model accuracy. In addition, in order to speed up the proposed quantization process during model training, we accelerate the quantization process with a parameterized probability estimation method and template-based derivation calculation. We evaluate our proposed algorithm on MNIST, CIFAR, ImageNet, IMDB movie review and THUCNews text data sets with numerical DNN models. The results demonstrate that our proposed quantization solution is more accurate and effective than the state-of-the-art approaches yet with more flexible bitwidth support. Moreover, the evaluation of our quantized models on Saliency Object Detection (SOD) tasks maintains comparable feature extraction quality with up to 16$\times$ weight size reduction.

cs.CV↗

arXiv2020

GPO: Global Plane Optimization for Fast and Accurate Monocular SLAM Initialization

Initialization is essential to monocular Simultaneous Localization and Mapping (SLAM) problems. This paper focuses on a novel initialization method for monocular SLAM based on planar features. The algorithm starts by homography estimation in a sliding window. It then proceeds to a global plane optimization (GPO) to obtain camera poses and the plane normal. 3D points can be recovered using planar constraints without triangulation. The proposed method fully exploits the plane information from multiple frames and avoids the ambiguities in homography decomposition. We validate our algorithm on the collected chessboard dataset against baseline implementations and present extensive analysis. Experimental results show that our method outperforms the fine-tuned baselines in both accuracy and real-time.

cs.RO↗

arXiv2020

HaoCL: Harnessing Large-scale Heterogeneous Processors Made Easy

The pervasive adoption of Deep Learning (DL) and Graph Processing (GP) makes it a de facto requirement to build large-scale clusters of heterogeneous accelerators including GPUs and FPGAs. The OpenCL programming framework can be used on the individual nodes of such clusters but is not intended for deployment in a distributed manner. Fortunately, the original OpenCL semantics naturally fit into the programming environment of heterogeneous clusters. In this paper, we propose a heterogeneity-aware OpenCL-like (HaoCL) programming framework to facilitate the programming of a wide range of scientific applications including DL and GP workloads on large-scale heterogeneous clusters. With HaoCL, existing applications can be directly deployed on heterogeneous clusters without any modifications to the original OpenCL source code and without awareness of the underlying hardware topologies and configurations. Our experiments show that HaoCL imposes a negligible overhead in a distributed environment, and provides near-linear speedups on standard benchmarks when computation or data size exceeds the capacity of a single node. The system design and the evaluations are presented in this demo paper.

cs.DC↗

arXiv2020

EDD: Efficient Differentiable DNN Architecture and Implementation Co-search for Embedded AI Solutions

High quality AI solutions require joint optimization of AI algorithms and their hardware implementations. In this work, we are the first to propose a fully simultaneous, efficient differentiable DNN architecture and implementation co-search (EDD) methodology. We formulate the co-search problem by fusing DNN search variables and hardware implementation variables into one solution space, and maximize both algorithm accuracy and hardware implementation quality. The formulation is differentiable with respect to the fused variables, so that gradient descent algorithm can be applied to greatly reduce the search time. The formulation is also applicable for various devices with different objectives. In the experiments, we demonstrate the effectiveness of our EDD methodology by searching for three representative DNNs, targeting low-latency GPU implementation and FPGA implementations with both recursive and pipelined architectures. Each model produced by EDD achieves similar accuracy as the best existing DNN models searched by neural architecture search (NAS) methods on ImageNet, but with superior performance obtained within 12 GPU-hour searches. Our DNN targeting GPU is 1.40x faster than the state-of-the-art solution reported in Proxyless, and our DNN targeting FPGA delivers 1.45x higher throughput than the state-of-the-art solution reported in DNNBuilder.

cs.LG↗

arXiv2020

TAG : Type Auxiliary Guiding for Code Comment Generation

Existing leading code comment generation approaches with the structure-to-sequence framework ignores the type information of the interpretation of the code, e.g., operator, string, etc. However, introducing the type information into the existing framework is non-trivial due to the hierarchical dependence among the type information. In order to address the issues above, we propose a Type Auxiliary Guiding encoder-decoder framework for the code comment generation task which considers the source code as an N-ary tree with type information associated with each node. Specifically, our framework is featured with a Type-associated Encoder and a Type-restricted Decoder which enables adaptive summarization of the source code. We further propose a hierarchical reinforcement learning method to resolve the training difficulties of our proposed framework. Extensive evaluations demonstrate the state-of-the-art performance of our framework with both the auto-evaluated metrics and case studies.

cs.CL↗

arXiv2020

An Extreme Ultraviolet Wave Associated with A Solar Filament Activation

Extreme ultraviolet (EUV) waves are impressive coronal propagating disturbances. They are closely associated with various eruptions, and can used for the global coronal seismology and the acceleration of solar energetic particles. Hence, the study of EUV waves plays an important role in solar eruptions and Space Weather. Here we present an EUV wave associated with a filament activation that did not evolve into any eruption. Due to the continuous magnetic flux emergence and cancellation around its one end, the filament rose with untwisting motion, and the filament mass flowed towards another end along the rising fields. Intriguingly, following the filament activation, an EUV wave formed with a fast constant speed ($\sim$500 km s$^{-1}$) ahead of the mass flow, and the overlying coronal loops expanded both in lateral and radial directions. Excluding the possibility of a remote flare and an absent coronal mass ejection, we suggest that the EUV wave was only closely associated with the filament activation. Furthermore, their intimate spacial and temporal relationship indicates that the EUV wave was likely directly triggered by the lateral expansion of overlying loops. We propose that the EUV wave can be interpreted as linear fast-mode wave, and the most vital key for the successful generation of the EUV wave is the impulsive early-phase lateral expansion of overlying loops that was driven by the activated filament mass flow without any eruption.

astro-ph.SR↗

arXiv2019

The Acceleration and Confinement of Energetic Electrons by a Termination Shock in a Magnetic Trap: An Explanation for Nonthermal Loop-top Sources during Solar Flares

Nonthermal loop-top sources in solar flares are the most prominent observational signature that suggests energy release and particle acceleration in the solar corona. Although several scenarios for particle acceleration have been proposed, the origin of the loop-top sources remains unclear. Here we present a model that combines a large-scale magnetohydrodynamic simulation of a two-ribbon flare with a particle acceleration and transport model for investigating electron acceleration by a fast-mode termination shock at the looptop. Our model provides spatially resolved electron distribution that evolves in response to the dynamic flare geometry. We find a concave-downward magnetic structure located below the flare termination shock, induced by the fast reconnection downflows. It acts as a magnetic trap to confine the electrons at the looptop for an extended period of time. The electrons are energized significantly as they cross the shock front, and eventually build up a power-law energy spectrum extending to hundreds of keV. We suggest that this particle acceleration and transport scenario driven by a flare termination shock is a viable interpretation for the observed nonthermal loop-top sources.

astro-ph.SR↗

arXiv2019

NAIS: Neural Architecture and Implementation Search and its Applications in Autonomous Driving

The rapidly growing demands for powerful AI algorithms in many application domains have motivated massive investment in both high-quality deep neural network (DNN) models and high-efficiency implementations. In this position paper, we argue that a simultaneous DNN/implementation co-design methodology, named Neural Architecture and Implementation Search (NAIS), deserves more research attention to boost the development productivity and efficiency of both DNN models and implementation optimization. We propose a stylized design methodology that can drastically cut down the search cost while preserving the quality of the end solution.As an illustration, we discuss this DNN/implementation methodology in the context of both FPGAs and GPUs. We take autonomous driving as a key use case as it is one of the most demanding areas for high quality AI algorithms and accelerators. We discuss how such a co-design methodology can impact the autonomous driving industry significantly. We identify several research opportunities in this exciting domain.

cs.LG↗

arXiv2019

Strong anomalous diffusion in two-state process with Lévy walk and Brownian motion

Strong anomalous diffusion phenomena are often observed in complex physical and biological systems, which are characterized by the nonlinear spectrum of exponents $qν(q)$ by measuring the absolute $q$-th moment $\langle |x|^q\rangle$. This paper investigates the strong anomalous diffusion behavior of a two-state process with Lévy walk and Brownian motion, which usually serves as an intermittent search process. The sojourn times in Lévy walk and Brownian phases are taken as power law distributions with exponents $α_+$ and $α_-$, respectively. Detailed scaling analyses are performed for the coexistence of three kinds of scalings in this system. Different from the pure Lévy walk, the phenomenon of strong anomalous diffusion can be observed for this two-state process even when the distribution exponent of Lévy walk phase satisfies $α_+<1$, provided that $α_-<α_+$. When $α_+<2$, the probability density function (PDF) in the central part becomes a combination of stretched Lévy distribution and Gaussian distribution due to the long sojourn time in Brownian phase, while the PDF in the tail part (in the ballistic scaling) is still dominated by the infinite density of Lévy walk.

cond-mat.stat-mech↗

arXiv2019

Langevin picture of Lévy walk in a constant force field

Lévy walk is a practical model and has wide applications in various fields. Here we focus on the effect of an external constant force on the Lévy walk with the exponent of the power-law distributed flight time $α\in(0,2)$. We add the term $Fη(s)$ ($η(s)$ is the Lévy noise) on a subordinated Langevin system to characterize such a constant force, being effective on the velocity process for all physical time after the subordination. We clearly show the effect of the constant force $F$ on this Langevin system and find this system is like the continuous limit of the collision model. The first moments of velocity processes for these two models are consistent. In particular, based on the velocity correlation function derived from our subordinated Langevin equation, we investigate more interesting statistical quantities, such as the ensemble- and time-averaged mean squared displacements. Under the influence of constant force, the diffusion of particles becomes faster. Finally, the super-ballistic diffusion and the non-ergodic behavior are verified by the simulations with different $α$.

cond-mat.stat-mech↗

arXiv2019

The Acceleration of Energetic Particles at Coronal Shocks and Emergence of a Double Power Law Feature in Particle Energy Spectra

We present numerical modelling of particle acceleration at coronal shocks propagating through a streamer-like magnetic field by solving the Parker transport equation with spatial diffusion both along and across the magnetic field. We show that the location on the shock where the high-energy particle intensity is the largest, depends on the energy of the particles and on time. The acceleration of particles to more than 100 MeV mainly occurs in the shock-streamer interaction region, due to perpendicular shock geometry and the trapping effect of closed magnetic fields. A comparison of the particle spectra to that in a radial magnetic field shows that the intensity at 100 MeV (200 MeV) is enhanced by more than one order (two orders) of magnitude. This indicates that the streamer-like magnetic field can be an important factor in producing large solar energetic particle events. We also show that the energy spectrum integrated over the simulation domain consists of two different power laws. Further analysis suggests that it may be a mixture of two distinct populations accelerated in the streamer and open field regions, where the acceleration rate differs substantially. Our calculations also show that the particle spectra are affected considerably by a number of parameters, such as the streamer tilt angle, particle spatial diffusion coefficient, and shock compression ratio. While the low-energy spectra agree well with standard diffusive shock acceleration theory, the break energy ranges from $\sim$1 MeV to $\sim$90 MeV and the high-energy spectra can extend to $\sim$1 GeV with a slope of $\sim$2-3.

astro-ph.SR↗

arXiv2019

Theory of relaxation dynamics for anomalous diffusion processes in harmonic potential

Optical tweezers setup is often used to probe the motion of individual tracer particle, which promotes the study of relaxation dynamics of a generic process confined in a harmonic potential. We uncover the dependence of ensemble- and time-averaged mean square displacements of confined processes on the velocity correlation function $C(t,t+τ)$ of the original process. With two different scaling forms of $C(t,t+τ)$ for small $τ$ and large $τ$, the stationary value and the relaxation behaviors can be obtained immediately. The gotten results are valid for a large amount of anomalous diffusion processes, including fractional Brownian motion, scaled Brownian motion, and the multi-scale Lévy walk with different exponents of running time distribution.

cond-mat.stat-mech↗