arXiv subjects

Pengfei Liu

Publications and source records attributed to Pengfei Liu.

269 records · Page 15Linked to original sources

arXiv2018

Meta Multi-Task Learning for Sequence Modeling

Semantic composition functions have been playing a pivotal role in neural representation learning of text sequences. In spite of their success, most existing models suffer from the underfitting problem: they use the same shared compositional function on all the positions in the sequence, thereby lacking expressive power due to incapacity to capture the richness of compositionality. Besides, the composition functions of different tasks are independent and learned from scratch. In this paper, we propose a new sharing scheme of composition function across multiple tasks. Specifically, we use a shared meta-network to capture the meta-knowledge of semantic composition and generate the parameters of the task-specific semantic composition models. We conduct extensive experiments on two types of tasks, text classification and sequence tagging, which demonstrate the benefits of our approach. Besides, we show that the shared meta-knowledge learned by our proposed model can be regarded as off-the-shelf knowledge and easily transferred to new tasks.

cs.AI↗

arXiv2018

Robust formation of topological Hall effect in MnGa/heavy metal bilayers

We have investigated the topological Hall effect (THE) in MnGa/Pt and MnGa/Ta bilayers induced by interfacial Dzyaloshinskii-Moriya interaction (DMI). The most evident THE signals have been found based on the MnGa films with small critical DMI energy constant Dc. The large topological portion of the Hall signal from the total Hall signal has been extracted in the whole temperature range from 5 to 300 K. These results open up the exploration of the DMI induced magnetic behavior based on the bulk perpendicular magnetic anisotropy materials for fundamental physics and magnetic storage technologies.

cond-mat.mtrl-sci↗

arXiv2017

A practical theoretical formalism for atomic multielectron processes: direct multiple ionization by a single Auger decay or by impact of a single electron or photon

Multiple electron processes occur widely in atoms, molecules, clusters, and condensed matters when they are interacting with energetic particles or intense laser fields. In this work, a unified and accurate theoretical formalism is proposed on the direct multielectron processes of atoms including the multiple Auger decay and multiple ionization by an impact of an energetic electron or a photon based on the atomic collision theory described by a correlated many-body Green's function. Such a practical treatment is made possible due to different coherence features of the particles (matter waves) in the initial and final states. We first explain how the coherence characteristics of the ejected continuum electrons is largely destructed, by taking the electron impact direct double ionization process as an example. This process is completely different from the single ionization where the complete interference can be maintained. The detailed expressions are obtained for the energy correlations among the continuum electrons and energy resolved differential and integral cross sections according to the separation of knock-out and shake-off mechanisms for the electron impact direct double ionization, direct double and triple Auger decay, and double and triple photoionization processes. The approach is applied to investigate the electron impact double ionization processes of C$^+$, N$^+$, and O$^+$, the direct double and triple Auger decay of the K-shell excited states of C$^{+}$ $1s2s^22p^2$ $^2D$ and $^2P$, and the double and triple photoionization of lithium. Comparisons with available experimental and theoretical results show that our proposed theoretical formalism is accurate and effective in treating the atomic multielectron processes.

physics.atom-ph↗

arXiv2017

Global regularity for a family of 3D models of the axisymmetric Navier-Stokes equations

We consider a family of 3D models for the axi-symmetric incompressible Navier-Stokes equations. The models are derived by changing the strength of the convection terms in the axisymmetric Navier-Stokes equations written using a set of transformed variables. We prove the global regularity of the family of models in the case that the strength of convection is slightly stronger than that of the original Navier-Stokes equations, which demonstrates the potential stabilizing effect of convection.

math.AP↗

arXiv2017

Dynamic Compositional Neural Networks over Tree Structure

Tree-structured neural networks have proven to be effective in learning semantic representations by exploiting syntactic information. In spite of their success, most existing models suffer from the underfitting problem: they recursively use the same shared compositional function throughout the whole compositional process and lack expressive power due to inability to capture the richness of compositionality. In this paper, we address this issue by introducing the dynamic compositional neural networks over tree structure (DC-TreeNN), in which the compositional function is dynamically generated by a meta network. The role of meta-network is to capture the metaknowledge across the different compositional rules and formulate them. Experimental results on two typical tasks show the effectiveness of the proposed models.

cs.CL↗

arXiv2017

Adversarial Multi-task Learning for Text Classification

Neural network models have shown their promising opportunities for multi-task learning, which focus on learning the shared layers to extract the common and task-invariant features. However, in most existing approaches, the extracted shared features are prone to be contaminated by task-specific features or the noise brought by other tasks. In this paper, we propose an adversarial multi-task learning framework, alleviating the shared and private latent feature spaces from interfering with each other. We conduct extensive experiments on 16 different text classification tasks, which demonstrates the benefits of our approach. Besides, we show that the shared knowledge learned by our proposed model can be regarded as off-the-shelf knowledge and easily transferred to new tasks. The datasets of all 16 tasks are publicly available at \url{http://nlp.fudan.edu.cn/data/}

cs.CL↗

arXiv2016

Deep Multi-Task Learning with Shared Memory

Neural network based models have achieved impressive results on various specific tasks. However, in previous works, most models are learned separately based on single-task supervised objectives, which often suffer from insufficient training data. In this paper, we propose two deep architectures which can be trained jointly on multiple related tasks. More specifically, we augment neural model with an external memory, which is shared by several tasks. Experiments on two groups of text classification tasks show that our proposed architectures can improve the performance of a task with the help of other related tasks.

cs.CL↗

arXiv2016

Modelling Interaction of Sentence Pair with coupled-LSTMs

Recently, there is rising interest in modelling the interactions of two sentences with deep neural networks. However, most of the existing methods encode two sequences with separate encoders, in which a sentence is encoded with little or no information from the other sentence. In this paper, we propose a deep architecture to model the strong interaction of sentence pair with two coupled-LSTMs. Specifically, we introduce two coupled ways to model the interdependences of two LSTMs, coupling the local contextualized interactions of two sentences. We then aggregate these interactions and use a dynamic pooling to select the most informative features. Experiments on two very large datasets demonstrate the efficacy of our proposed architecture and its superiority to state-of-the-art methods.

cs.CL↗

arXiv2016

Recurrent Neural Network for Text Classification with Multi-Task Learning

Neural network based methods have obtained great progress on a variety of natural language processing tasks. However, in most previous works, the models are learned based on single-task supervised objectives, which often suffer from insufficient training data. In this paper, we use the multi-task learning framework to jointly learn across multiple related tasks. Based on recurrent neural network, we propose three different mechanisms of sharing information to model text with task-specific and shared layers. The entire network is trained jointly on all these tasks. Experiments on four benchmark text classification tasks show that our proposed models can improve the performance of a task with the help of other related tasks.

cs.CL↗

arXiv2015

Optimal Local Multi-scale Basis Functions for Linear Elliptic Equations with Rough Coefficient

This paper addresses a multi-scale finite element method for second order linear elliptic equations with arbitrarily rough coefficient. We propose a local oversampling method to construct basis functions that have optimal local approximation property. Our methodology is based on the compactness of the solution operator restricted on local regions of the spatial domain, and does not depend on any scale-separation or periodicity assumption of the coefficient. We focus on a special type of basis functions that are harmonic on each element and have optimal approximation property. We first reduce our problem to approximating the trace of the solution space on each edge of the underlying mesh, and then achieve this goal through the singular value decomposition of an oversampling operator. Rigorous error estimates can be obtained through thresholding in constructing the basis functions. Numerical results for several problems with multiple spatial scales and high contrast inclusions are presented to demonstrate the compactness of the local solution space and the capacity of our method in identifying and exploiting this compact structure to achieve computational savings.

math.NA↗

arXiv2015

Local-set-based Graph Signal Reconstruction

Signal processing on graph is attracting more and more attentions. For a graph signal in the low-frequency subspace, the missing data associated with unsampled vertices can be reconstructed through the sampled data by exploiting the smoothness of the graph signal. In this paper, the concept of local set is introduced and two local-set-based iterative methods are proposed to reconstruct bandlimited graph signal from sampled data. In each iteration, one of the proposed methods reweights the sampled residuals for different vertices, while the other propagates the sampled residuals in their respective local sets. These algorithms are built on frame theory and the concept of local sets, based on which several frames and contraction operators are proposed. We then prove that the reconstruction methods converge to the original signal under certain conditions and demonstrate the new methods lead to a significantly faster convergence compared with the baseline method. Furthermore, the correspondence between graph signal sampling and time-domain irregular sampling is analyzed comprehensively, which may be helpful to future works on graph signals. Computer simulations are conducted. The experimental results demonstrate the effectiveness of the reconstruction methods in various sampling geometries, imprecise priori knowledge of cutoff frequency, and noisy scenarios.

cs.IT↗

arXiv2015

Self-similar Singularity of a 1D Model for the 3D Axisymmetric Euler Equations

We investigate the self-similar singularity of a 1D model for the 3D axisymmetric Euler equations, which is motivated by a particular singularity formation scenario observed in numerical computation. We prove the existence of a discrete family of self-similar profiles for this model and analyze their far-field properties. The self-similar profiles we find agree with direct simulation of the model and seem to have some stability.

math.AP↗

arXiv2014

A Heterogeneous Stochastic FEM Framework for Elliptic PDEs

We introduce a new concept of sparsity for the stochastic elliptic operator $-{\rm div}\left(a(x,ω)\nabla(\cdot)\right)$, which reflects the compactness of its inverse operator in the stochastic direction and allows for spatially heterogeneous stochastic structure. This new concept of sparsity motivates a heterogeneous stochastic finite element method ({\bf HSFEM}) framework for linear elliptic equations, which discretizes the equations using the heterogeneous coupling of spatial basis with local stochastic basis to exploit the local stochastic structure of the solution space. We also provide a sampling method to construct the local stochastic basis for this framework using the randomized range finding techniques. The resulting HSFEM involves two stages and suits the multi-query setting: in the offline stage, the local stochastic structure of the solution space is identified; in the online stage, the equation can be efficiently solved for multiple forcing functions. An online error estimation and correction procedure through Monte Carlo sampling is given. Numerical results for several problems with high dimensional stochastic input are presented to demonstrate the efficiency of the HSFEM in the online stage.

math.NA↗

arXiv2013

A superradiant laser based on two-photon Raman transition of caesium atoms

We propose a superradiant laser based on two-photon Raman transition of caesium-133 atoms which collectively emit photons on an ultra narrow transition into the mode of a low Q resonator known as optical bad-cavity regime. The spin-spin correlation which characterizes the collective effect is demonstrated. We theoretically predict that the optical radiation has an extremely narrow linewidth in the 98 (1) *10-2 mHz range, smaller than the transition itself due to collective effects, and a power level of 7 (1)*10-10 W is possible, which can provide a possible new way to realize an optical clock with a millihertz linewidth.

physics.optics↗

arXiv2013

Edge Balance Ratio: Power Law from Vertices to Edges in Directed Complex Network

Power law distribution is common in real-world networks including online social networks. Many studies on complex networks focus on the characteristics of vertices, which are always proved to follow the power law. However, few researches have been done on edges in directed networks. In this paper, edge balance ratio is firstly proposed to measure the balance property of edges in directed networks. Based on edge balance ratio, balance profile and positivity are put forward to describe the balance level of the whole network. Then the distribution of edge balance ratio is theoretically analyzed. In a directed network whose vertex in-degree follows the power law with scaling exponent $γ$, it is proved that the edge balance ratio follows a piecewise power law, with the scaling exponent of each section linearly dependents on $γ$. The theoretical analysis is verified by numerical simulations. Moreover, the theoretical analysis is confirmed by statistics of real-world online social networks, including Twitter network with 35 million users and Sina Weibo network with 110 million users.

cs.SI↗

arXiv2012

Follow Whom? Chinese Users Have Different Choice

Sina Weibo, which was launched in 2009, is the most popular Chinese micro-blogging service. It has been reported that Sina Weibo has more than 400 million registered users by the end of the third quarter in 2012. Sina Weibo and Twitter have a lot in common, however, in terms of the following preference, Sina Weibo users, most of whom are Chinese, behave differently compared with those of Twitter. This work is based on a data set of Sina Weibo which contains 80.8 million users' profiles and 7.2 billion relations and a large data set of Twitter. Firstly some basic features of Sina Weibo and Twitter are analyzed such as degree and activeness distribution, correlation between degree and activeness, and the degree of separation. Then the following preference is investigated by studying the assortative mixing, friend similarities, following distribution, edge balance ratio, and ranking correlation, where edge balance ratio is newly proposed to measure balance property of graphs. It is found that Sina Weibo has a lower reciprocity rate, more positive balanced relations and is more disassortative. Coinciding with Asian traditional culture, the following preference of Sina Weibo users is more concentrated and hierarchical: they are more likely to follow people at higher or the same social levels and less likely to follow people lower than themselves. In contrast, the same kind of following preference is weaker in Twitter. Twitter users are open as they follow people from levels, which accords with its global characteristic and the prevalence of western civilization. The message forwarding behavior is studied by displaying the propagation levels, delays, and critical users. The following preference derives from not only the usage habits but also underlying reasons such as personalities and social moralities that is worthy of future research.

cs.SI↗

arXiv2012

Thermal modelling for endocardiac radiofrequency ablation: comparison of hyperbolic bioheat equation and Pennes bioheat equation with finite element method

The objectives of this study are to model the endocardiac radiofrequency (RF) ablation procedure and to employ the Hyperbolic Bioheat Equation (HBE), which takes the thermal wave behaviour into account, comparing the results with those obtained using the common Pennes Bioheat Equation (BE) method. A complex model is created to cover particular endocardiac physical and geometry environment. Finite Element Method (FEM) is adopted to study the model with both BE and HBE methods. Different convection coefficients and voltages are applied to simulate different conditions. Lesion size, max temperature and specified position temperature are selected as criteria to evaluate the simulated results. The study found that during ablation, the lesion size difference ratio can reach 20% in some periods. The difference is obvious and cannot be neglected.

math.AP↗