arXiv subjects

Wei Xu

Publications and source records attributed to Wei Xu.

At least 577 records · Page 32Linked to original sources

arXiv2018

Zero-Shot Transfer VQA Dataset

Acquiring a large vocabulary is an important aspect of human intelligence. Onecommon approach for human to populating vocabulary is to learn words duringreading or listening, and then use them in writing or speaking. This ability totransfer from input to output is natural for human, but it is difficult for machines.Human spontaneously performs this knowledge transfer in complicated multimodaltasks, such as Visual Question Answering (VQA). In order to approach human-levelArtificial Intelligence, we hope to equip machines with such ability. Therefore, toaccelerate this research, we propose a newzero-shot transfer VQA(ZST-VQA)dataset by reorganizing the existing VQA v1.0 dataset in the way that duringtraining, some words appear only in one module (i.e. questions) but not in theother (i.e. answers). In this setting, an intelligent model should understand andlearn the concepts from one module (i.e. questions), and at test time, transfer themto the other (i.e. predict the concepts as answers). We conduct evaluation on thisnew dataset using three existing state-of-the-art VQA neural models. Experimentalresults show a significant drop in performance on this dataset, indicating existingmethods do not address the zero-shot transfer problem. Besides, our analysis findsthat this may be caused by the implicit bias learned during training.

cs.AI↗

arXiv2018

A Word-Complexity Lexicon and A Neural Readability Ranking Model for Lexical Simplification

Current lexical simplification approaches rely heavily on heuristics and corpus level features that do not always align with human judgment. We create a human-rated word-complexity lexicon of 15,000 English words and propose a novel neural readability ranking model with a Gaussian-based feature vectorization layer that utilizes these human ratings to measure the complexity of any given word or phrase. Our model performs better than the state-of-the-art systems for different lexical simplification tasks and evaluation datasets. Additionally, we also produce SimplePPDB++, a lexical resource of over 10 million simplifying paraphrase rules, by applying our model to the Paraphrase Database (PPDB).

cs.CL↗

arXiv2018

Joint Unsupervised Learning of Optical Flow and Depth by Watching Stereo Videos

Learning depth and optical flow via deep neural networks by watching videos has made significant progress recently. In this paper, we jointly solve the two tasks by exploiting the underlying geometric rules within stereo videos. Specifically, given two consecutive stereo image pairs from a video, we first estimate depth, camera ego-motion and optical flow from three neural networks. Then the whole scene is decomposed into moving foreground and static background by compar- ing the estimated optical flow and rigid flow derived from the depth and ego-motion. We propose a novel consistency loss to let the optical flow learn from the more accurate rigid flow in static regions. We also design a rigid alignment module which helps refine ego-motion estimation by using the estimated depth and optical flow. Experiments on the KITTI dataset show that our results significantly outperform other state-of- the-art algorithms. Source codes can be found at https: //github.com/baidu-research/UnDepthflow

cs.CV↗

arXiv2018

Performance Analysis of Multi-Cell Millimeter Wave Massive MIMO Networks with Low-Precision ADCs

In this paper, we investigate a multi-cell millimeter wave (mmWave) massive multiple-input multiple-output (MIMO) network with low-precision analog-to-digital converters (ADCs) at the base station (BS). Each cell serves multiple users and each user is equipped with multiple antennas but driven by a single RF chain. We first introduce a channel estimation strategy for the mmWave massive MIMO network and analyze the achievable rate with imperfect channel state information. Then, we derive an insightful lower bound for the achievable rate, which becomes tight with a growing number of users. The bound clearly demonstrates the impacts of the number of antennas and the ADC precision, especially for a single-cell mmWave network at low signal-to-noise ratio (SNR). It characterizes the tradeoff among various system parameters. Our analytical results are finally confirmed by extensive computer simulations.

cs.IT↗

arXiv2018

Wideband mmWave Channel Estimation for Hybrid Massive MIMO with Low-Precision ADCs

In this article, we investigate channel estimation for wideband millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) under hybrid architecture with lowprecision analog-to-digital converters (ADCs). To design channel estimation for the hybrid structure, both analog processing components and frequency-selective digital combiners need to be optimized. The proposed channel estimator follows the typical linear-minimum-mean-square-error (LMMSE) structure and applies for an arbitrary channel model. Moreover, for sparsity channels as in mmWave, the proposed estimator performs more efficiently by incorporating orthogonal matching pursuit (OMP) to mitigate quantization noise caused by low-precision ADCs. Consequently, the proposed estimator outperforms conventional ones as demonstrated by computer simulation results.

eess.SP↗

arXiv2018

Improving Wireless Physical Layer Security via D2D Communication

This paper investigates the physical layer security issue of a device-to-device (D2D) underlaid cellular system with a multi-antenna base station (BS) and a multi-antenna eavesdropper. To investigate the potential of D2D communication in improving network security, the conventional network without D2D users (DUs) is first considered. It is shown that the problem of maximizing the sum secrecy rate (SR) of cellular users (CUs) for this special case can be transformed to an assignment problem and optimally solved. Then, a D2D underlaid network is considered. Since the joint optimization of resource block (RB) allocation, CU-DU matching and power control is a mixed integer programming, the problem is difficult to handle. Hence, the RB assignment process is first conducted by ignoring D2D communication, and an iterative algorithm is then proposed to solve the remaining problem. Simulation results show that the sum SR of CUs can be greatly increased by D2D communication, and compared with the existing schemes, a better secrecy performance can be obtained by the proposed algorithms.

cs.IT↗

arXiv2018

Guided Feature Transformation (GFT): A Neural Language Grounding Module for Embodied Agents

Recently there has been a rising interest in training agents, embodied in virtual environments, to perform language-directed tasks by deep reinforcement learning. In this paper, we propose a simple but effective neural language grounding module for embodied agents that can be trained end to end from scratch taking raw pixels, unstructured linguistic commands, and sparse rewards as the inputs. We model the language grounding process as a language-guided transformation of visual features, where latent sentence embeddings are used as the transformation matrices. In several language-directed navigation tasks that feature challenging partial observability and require simple reasoning, our module significantly outperforms the state of the art. We also release XWorld3D, an easy-to-customize 3D environment that can potentially be modified to evaluate a variety of embodied agents.

cs.AI↗

arXiv2018

Scaling Nakamoto Consensus to Thousands of Transactions per Second

This paper presents Conflux, a fast, scalable and decentralized blockchain system that optimistically process concurrent blocks without discarding any as forks. The Conflux consensus protocol represents relationships between blocks as a direct acyclic graph and achieves consensus on a total order of the blocks. Conflux then, from the block order, deterministically derives a transaction total order as the blockchain ledger. We evaluated Conflux on Amazon EC2 clusters with up to 20k full nodes. Conflux achieves a transaction throughput of 5.76GB/h while confirming transactions in 4.5-7.4 minutes. The throughput is equivalent to 6400 transactions per second for typical Bitcoin transactions. Our results also indicate that when running Conflux, the consensus protocol is no longer the throughput bottleneck. The bottleneck is instead at the processing capability of individual nodes.

cs.DC↗

arXiv2018

Simulation of Heterogeneous Atom Probe Tip Shapes Evolution during Field Evaporation Using a Level Set Method and Different Evaporation Models

In atom probe tomography (APT), accurate reconstruction of the spatial positions of field evaporated ions from measured detector patterns depends upon a correct understanding of the dynamic tip shape evolution and evaporation laws of component atoms. Artifacts in APT reconstructions of heterogeneous materials can be attributed to the assumption of homogeneous evaporation of all the elements in the material in addition to the assumption of a steady state hemispherical dynamic tip shape evolution. A level set method based specimen shape evolution model is developed in this study to simulate the evaporation of synthetic layered-structured APT tips. The simulation results of the shape evolution by the level set model qualitatively agree with the finite element method and the literature data using the finite difference method. The asymmetric evolving shape predicted by the level set model demonstrates the complex evaporation behavior of heterogeneous tip and the interface curvature can potentially lead to the artifacts in the APT reconstruction of such materials. Compared with other APT simulation methods, the new method provides smoother interface representation with the aid of the intrinsic sub-grid accuracy. Two evaporation models (linear and exponential evaporation laws) are implemented in the level set simulations and the effect of evaporation laws on the tip shape evolution is also presented.

physics.comp-ph↗

arXiv2018

High-Performance Multi-Mode Ptychography Reconstruction on Distributed GPUs

Ptychography is an emerging imaging technique that is able to provide wavelength-limited spatial resolution from specimen with extended lateral dimensions. As a scanning microscopy method, a typical two-dimensional image requires a number of data frames. As a diffraction-based imaging technique, the real-space image has to be recovered through iterative reconstruction algorithms. Due to these two inherent aspects, a ptychographic reconstruction is generally a computation-intensive and time-consuming process, which limits the throughput of this method. We report an accelerated version of the multi-mode difference map algorithm for ptychography reconstruction using multiple distributed GPUs. This approach leverages available scientific computing packages in Python, including mpi4py and PyCUDA, with the core computation functions implemented in CUDA C. We find that interestingly even with MPI collective communications, the weak scaling in the number of GPU nodes can still remain nearly constant. Most importantly, for realistic diffraction measurements, we observe a speedup ranging from a factor of $10$ to $10^3$ depending on the data size, which reduces the reconstruction time remarkably from hours to typically about 1 minute and is thus critical for real-time data processing and visualization.

physics.comp-ph↗

arXiv2018

A Framework on Hybrid MIMO Transceiver Design based on Matrix-Monotonic Optimization

Hybrid transceiver can strike a balance between complexity and performance of multiple-input multiple-output (MIMO) systems. In this paper, we develop a unified framework on hybrid MIMO transceiver design using matrix-monotonic optimization. The proposed framework addresses general hybrid transceiver design, rather than just limiting to certain high frequency bands, such as millimeter wave (mmWave) or terahertz bands or relying on the sparsity of some specific wireless channels. In the proposed framework, analog and digital parts of a transceiver, either linear or nonlinear, are jointly optimized. Based on matrix-monotonic optimization, we demonstrate that the combination of the optimal analog precoders and processors are equivalent to eigenchannel selection for various optimal hybrid MIMO transceivers. From the optimal structure, several effective algorithms are derived to compute the analog transceivers under unit modulus constraints. Furthermore, in order to reduce computation complexity, a simple random algorithm is introduced for analog transceiver optimization. Once the analog part of a transceiver is determined, the closed-form digital part can be obtained. Numerical results verify the advantages of the proposed design.

cs.IT↗

arXiv2018

An Enhanced SCMA Detector Enabled by Deep Neural Network

In this paper, we propose a learning approach for sparse code multiple access (SCMA) signal detection by using a deep neural network via unfolding the procedure of message passing algorithm (MPA). The MPA can be converted to a sparsely connected neural network if we treat the weights as the parameters of a neural network. The neural network can be trained off-line and then deployed for online detection. By further refining the network weights corresponding to the edges of a factor graph, the proposed method achieves a better performance. Moreover, the deep neural network based detection is a computationally efficient since highly paralleled computations in the network are enabled in emerging Artificial Intelligence (AI) chips.

cs.IT↗

arXiv2018

Neural Network Models for Paraphrase Identification, Semantic Textual Similarity, Natural Language Inference, and Question Answering

In this paper, we analyze several neural network designs (and their variations) for sentence pair modeling and compare their performance extensively across eight datasets, including paraphrase identification, semantic textual similarity, natural language inference, and question answering tasks. Although most of these models have claimed state-of-the-art performance, the original papers often reported on only one or two selected datasets. We provide a systematic study and show that (i) encoding contextual information by LSTM and inter-sentence interactions are critical, (ii) Tree-LSTM does not help as much as previously claimed but surprisingly improves performance on Twitter datasets, (iii) the Enhanced Sequential Inference Model is the best so far for larger datasets, while the Pairwise Word Interaction Model achieves the best performance when less data is available. We release our implementations as an open-source toolkit.

cs.CL↗

arXiv2018

Every Pixel Counts: Unsupervised Geometry Learning with Holistic 3D Motion Understanding

Learning to estimate 3D geometry in a single image by watching unlabeled videos via deep convolutional network has made significant process recently. Current state-of-the-art (SOTA) methods, are based on the learning framework of rigid structure-from-motion, where only 3D camera ego motion is modeled for geometry estimation.However, moving objects also exist in many videos, e.g. moving cars in a street scene. In this paper, we tackle such motion by additionally incorporating per-pixel 3D object motion into the learning framework, which provides holistic 3D scene flow understanding and helps single image geometry estimation. Specifically, given two consecutive frames from a video, we adopt a motion network to predict their relative 3D camera pose and a segmentation mask distinguishing moving objects and rigid background. An optical flow network is used to estimate dense 2D per-pixel correspondence. A single image depth network predicts depth maps for both images. The four types of information, i.e. 2D flow, camera pose, segment mask and depth maps, are integrated into a differentiable holistic 3D motion parser (HMP), where per-pixel 3D motion for rigid background and moving objects are recovered. We design various losses w.r.t. the two types of 3D motions for training the depth and motion networks, yielding further error reduction for estimated geometry. Finally, in order to solve the 3D motion confusion from monocular videos, we combine stereo images into joint training. Experiments on KITTI 2015 dataset show that our estimated geometry, 3D motion and moving object masks, not only are constrained to be consistent, but also significantly outperforms other SOTA algorithms, demonstrating the benefits of our approach.

cs.CV↗

arXiv2018

Energy-Efficient Multi-View Video Transmission with View Synthesis-Enabled Multicast

Multi-view videos (MVVs) provide immersive viewing experience, at the cost of heavy load to wireless networks. Except for further improving viewing experience, view synthesis can create multicast opportunities for efficient transmission of MVVs in multiuser wireless networks, which has not been recognized in existing literature. In this paper, we would like to exploit view synthesis-enabled multicast opportunities for energy-efficient MVV transmission in a multiuser wireless network. Specifically, we first establish a mathematical model to characterize the impact of view synthesis on multicast opportunities and energy consumption. Then, we consider the optimization of view selection, transmission time and power allocation to minimize the weighted sum energy consumption for view transmission and synthesis, which is a challenging mixed discrete-continuous optimization problem. We propose an algorithm to obtain an optimal solution with reduced computational complexity by exploiting optimality properties. To further reduce computational complexity, we also propose two low-complexity algorithms to obtain two suboptimal solutions, based on continuous relaxation and Difference of Convex (DC) programming, respectively. Finally, numerical results demonstrate the advantage of the proposed solutions.

cs.IT↗

arXiv2018

Interactive Grounded Language Acquisition and Generalization in a 2D World

We build a virtual agent for learning language in a 2D maze-like world. The agent sees images of the surrounding environment, listens to a virtual teacher, and takes actions to receive rewards. It interactively learns the teacher's language from scratch based on two language use cases: sentence-directed navigation and question answering. It learns simultaneously the visual representations of the world, the language, and the action control. By disentangling language grounding from other computational routines and sharing a concept detection function between language grounding and prediction, the agent reliably interpolates and extrapolates to interpret sentences that contain new word combinations or new words missing from training sentences. The new words are transferred from the answers of language prediction. Such a language ability is trained and evaluated on a population of over 1.6 million distinct sentences consisting of 119 object words, 8 color words, 9 spatial-relation words, and 50 grammatical words. The proposed model significantly outperforms five comparison methods for interpreting zero-shot sentences. In addition, we demonstrate human-interpretable intermediate outputs of the model in the appendix.

cs.CL↗

arXiv2018

Ferromagnetism in graphene traced to an antisymmetric orbital combination of involved electronic states

We reveal that the origin of ferromagnetism caused by $sp$ electrons in graphene with vacancies can be traced to electrons partially filling $sp^{2*}$-antibonding and $p_z^*$-nonbonding states, which are induced by the vacancies and appear near the Fermi level. Because the spatial wavefunctions of the both states are composed of atomic orbitals in an antisymmetric configuration, their spin wavefunctions should be symmetric according to the electron exchange antisymmetric principle, leading to electrons partially filling these states in spin polarization. Since this $p_z^*$ state originates not from interactions between the atoms but from the unpaired $p_z$ orbitals due to the removal of $p_z$ orbitals on the minority sublattice, the $p_z^*$ state is constrained, distributed on the atoms of the majority sublattice, and decays gradually from the vacancy as $\sim$ $1/r$. According to these characteristics, we concluded that the $p_z^*$ state plays a critical role in magnetic ordering in graphene with vacancies. If the vacancy concentration in graphene is large enough to cause the decay-length regions to overlap, constraining the $p_z^*$ orbital components as little as possible on the minority sublattice atoms in the overlap regions results in the vacancy-induced $p_z^*$ states being coherent. The coherent process in the overlap region leads to the wavefunctions in all the involved regions antisymmetrized, consequently causing ferromagnetism according to the electron exchange antisymmetric principle. This unusual mechanism concerned with the origin of $sp$-electron magnetism and magnetic ordering has never before been reported and is distinctly different from conventional mechanisms.

cond-mat.mtrl-sci↗

arXiv2018

Origin of $sp$-electron magnetism in Graphitic Carbon Nitride

Based on first principles calculations, this study reveals that magnetism in otherwise non-magnetic materials can originate from the partial occupation of antibonding states. Since the antibonding wavefunctions are spatially antisymmetric, the spin wavefunctions should be symmteric according to the exchange antisymmetric principle of quantum mechanics. We demonstrate that this phenomenon can be observed in a graphitic carbon nitride material, $g$-C$_4$N$_3$, which can be experimentally synthesized and seen as a honeycomb structure with a vacancy. Three dangling bonds of N atoms pointing to the vacancy site interact with each other to form one bonding and two antibonding states. As the two antibonding states are near the Fermi level, and electrons should partially occupy the antibonding states in spin polarization, this leads to 1~$μ_B$ magnetic moment.

cond-mat.mtrl-sci↗