We address the problem of serving Deep Neural Networks (DNNs) efficiently from a cluster GPUs. In order to realize promise very low-cost processing made by accelerators such as GPUs, it is essential run them at sustained high utilization. Doing so requires cluster-scale resource management that performs detailed scheduling reasoning about groups DNN invocations need be co-scheduled, and moving conventional whole-DNN execution model executing fragments DNNs. Nexus fully implemented system...

In this paper, we discuss an unsupervised deep learning (DL) method for solving time domain electromagnetic simulations. Compared to the conventional approach, our encodes initial conditions, boundary conditions as well Maxwell's equations constraints when training network, turning simulation problem into optimization process. High prediction accuracy of fields, without discretization or interpolation in space time, can be achieved with limited number layers and neurons each layer neural...

Energy consumption of mobile apps has become an important consideration as the underlying devices are constrained by battery capacity. Display represents a significant portion app's energy consumption. However, developers lack techniques to identify user interfaces in their for which needs be improved. In this paper, we present technique detecting display hotspots - app whose is greater than optimal. Our leverages power modeling and automated transformation detect these prioritize them...

In this article, a new scheme based on the supervised descent method (SDM) for solving directional electromagnetic logging-while-drilling (LWD) inverse problems is proposed. The SDM provides us perspective to combine classical gradient-based inversion and machine-learning-based schemes. It iteratively learns set of directions in offline training process, where model generated advance according prior information, then updates models with learned as well data residuals prediction stage,...

To effectively overcome the cycle-skipping issue in full-waveform inversion (FWI), we have developed a deep neural network (DNN) approach to predict absent low-frequency (LF) components by exploiting hidden physical relation connecting LF and high-frequency (HF) data. efficiently solve this challenging nonlinear regression problem, two novel strategies are proposed design DNN architecture optimize learning process: (1) dual data feed structure (2) progressive transfer learning. With...

In situ testing techniques have become an important means of ensuring the reliability embedded systems after they are deployed in field. However, these do not help testers optimize energy consumption their test suites, which can needlessly waste limited battery power systems. this work, we extend prior for suite minimization such a way as to allow generate energy-efficient, minimized suites with only minimal modifications existing work flow. We perform extensive empirical evaluation our...

In this paper, we will explore the possibility of synthesizing low-frequency data from high-frequency data. The synthesized are used to improve full-waveform inversion (FWI). Unlike all previously methods, best our knowledge, is first attempt utilize a driven approach solve problem. We propose learn low wavenumber information in FWI via Deep Inception based Convolutional Networks. Once deep learning network sufficiently trained, can be predicted with high accuracy on completely different...

Consumer cloud storage (CCS) services have become popular among users for storing and synchronizing files via apps installed on their devices. A single CCS, however, has intrinsic limitations networking performance, service reliability, data security. To overcome these limitations, we present UniDrive, a CCS app that synergizes multiple CCSs (multi-cloud) by using only few simple public RESTful Web APIs. UniDrive follows server-less, client-centric design, in which synchronization logic is...

The network communications between the cloud and client have become weak link for global services that aim to provide low latency their clients. In this paper, we first characterize WAN from viewpoint of a large provider Azure, whose edges serve hundreds billions TCP connections day across locations worldwide. particular, focus on instances degradation design tool, BlameIt, enables operators localize cause (i.e., faulty AS) such degradation. BlameIt uses passive diagnosis, using measurements...

In this communication, a trainable theory-guided recurrent neural network (RNN) equivalent to the finite-difference-time-domain (FDTD) method is exploited formulate electromagnetic propagation, solve Maxwell's equations, and inverse problem on differentiable programming platform Pytorch. For forward modeling, computation efficiency substantially improved compared conventional FDTD implemented MATLAB. Gradient becomes more precise faster than traditional finite difference benefiting from...

Deep learning leverages multi-layer neural networks architecture and demonstrates superb power in many machine applications. The deep denoising autoencoder technique extracts better coherent features from the seismic data. allows us to automatically extract low-dimensional high dimensional feature space a non-linear, data-driven, unsupervised way. A properly trained takes partially corrupted input recovers original undistorted input. In this paper, novel built upon residual network is...

In many geophysical machine learning applications, labeled data are either too expensive or impossible to collect. this work, we developed a physics-guided self-supervised method predict the absent low frequency (LF) components in acquired seismic overcome cycle-skipping issue full waveform inversion (FWI). Similar Natural Language Processing (NLP), label-free consists of two stages: pretext task and downstream task. For task, proposed algorithms, pseudo-LF (PLF) sparse inversion-based...

The low-frequency seismic data provide crucial information for guiding the full-waveform inversion (FWI), especially when strong reflectors exist in velocity model. However, hardware limitations make it difficult to acquire data. To overcome nonlinearity and ill-posedness caused by absence of data, we develop an efficient progressive transfer learning algorithm extrapolation. proposed method combines FWI, sparsity-promoted bandwidth-extension (BWE) algorithm, physics-guided data-driven deep...

Genetic architecture of asthma remains obscure. This study aimed to investigate whether the genetic polymorphisms CDHR3 (rs6967330), GSDMB (rs2305480), IL33 rs928413, RAD50 (rs6871536) and IL1RL1 (rs1558641) are associated with development atopic in Chinese population. Genotype allele frequencies were compared between 516 patients 552 controls by Chi-square test. Patients found have significantly higher G rs928413 C rs6871536 (9.5% vs 6.2%, P = 0.004 for rs928413; 26.1% 19.9%, < 0.001...

Summary The energy consumption of mobile apps has become an important consideration for developers as the underlying devices are constrained by battery capacity. Display represents a significant portion app's consumption—up to 60% total consumption. However, lack techniques identify user interfaces in their which needs be improved. This paper presents technique detecting display hotspots—user app whose is greater than optimal. leverages power modeling and automated transformation detect...

To effectively mitigate the cycle-skipping phenomenon in full waveform inversion (FWI), we developed an inception-based deep learning neural network to reconstruct absent low frequency data by exploiting subsurface wavenumber information buried acquired high data. Two unique features of our for reconstruction are 1) Dual Data Feed; 2) Progressive Transfer Learning. With Feed feature, not only data, but also corresponding Beat Tone fed into network, significantly reducing complexity and...

In this study, we discuss an unsupervised deep learning approach for time-domain electromagnetic simulations. Our method, based on physics informed neural network, encodes initial conditions, boundary conditions as well governing equations the constraints when training turning simulation problem into optimization process. High prediction accuracy of fields, without discretization or interpolation in space time, can be achieved with limited numbers layers and neurons each layer network. We...

We report the synthesis of poly(acrylamide-co-acrylic acid)/sodium carboxy methyl cellulose (PAMAA/CMC-Na) hydrogels, and subsequent fabrication dual-network polymer hydrogels (PAMAA/CMC-Na/Fe) using as-prepared via salt solution (FeCl3) immersion method. The created exhibit anti-swelling properties, frost resistance, high conductivity, good mechanical performance. hydrogel swells sightly when immersed in (pH = 2~11). With increase nAA:nAM, modulus elasticity experiences a rise from 1.1 to...

Various network optimization and management goals in cloud data centers can be achieved by re-mapping VMs to the substrate servers. Live VM migration is used implement such moving from initial servers target ones. For efficiency usability considerations, these tasks are expected completed as soon possible, which planning multiple migrated simultaneously. However, available resources of computation bandwidth limits number that at same time. Besides, change with progress process, makes a...

Full-waveform inversion (FWI) plays a significant role in producing high-resolution subsurface imaging seismic prospecting and ground penetrating radar (GPR). However, FWI faces various challenges practice. For example, the lack of low-frequency information due to acquisition limitations will make prone falling local minimum. In this project, deep learning-based approach is proposed extrapolate data. Specifically, we propose robust progressive learning (RPL) algorithm that combines...