Searcharxiv⌕ Search

arXiv subjects

Bo Chen

Publications and source records attributed to Bo Chen.

At least 415 records · Page 23Linked to original sources
arXiv2017

Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference

The rising popularity of intelligent mobile devices and the daunting computational cost of deep learning-based models call for efficient and accurate on-device inference schemes. We propose a quantization scheme that allows inference to be carried out using integer-only arithmetic, which can be implemented more efficiently than floating point inference on commonly available integer-only hardware. We also co-design a training procedure to preserve end-to-end model accuracy post quantization. As a result, the proposed quantization scheme improves the tradeoff between accuracy and on-device latency. The improvements are significant even on MobileNets, a model family known for run-time efficiency, and are demonstrated in ImageNet classification and COCO detection on popular CPUs.

cs.LG↗
arXiv2017

Approaching the Type-II Dirac Point and Concomitant Superconductivity in Pt-doping Stabilized Metastable 1T-phase IrTe2

Topological semimetal is a topic of general interest in material science. Recently, a new kind of topological semimetal called type-II Dirac semimetal with tilted Dirac cones is discovered in PtSe2 family. However, the further investigation is hindered due to the huge energy difference from Dirac points to Fermi level and the irrelevant conducting pockets at Fermi surface. Here we characterize the optimized type-II Dirac dispersions in a metastable 1T phase of IrTe2. Our strategy of Pt doping protects the metastable 1T phase in low temperature and tunes the Fermi level to the Dirac point. As demonstrated by angle-resolved photoemission spectra and first principle calculations, the Fermi surface of Ir1-xPtxTe2 is formed by only a single band with type-II Dirac cone which is tilted strongly along kz momentum direction. Interesting superconductivity is observed in samples for Dirac point close to Fermi level and even survives when Fermi level aligns with the Dirac point as finite density of states created by the tilted cone dispersion. This advantage offers opportunities for possible topological superconductivity and versatile Majorana devices in type-II Dirac semimetals.

cond-mat.mtrl-sci↗
arXiv2017

Error Estimates for Sparse Optimal Control Problems by Piecewise Linear Finite Element Approximation

Optimization problems with $L^1$-control cost functional subject to an elliptic partial differential equation (PDE) are considered. However, different from the finite dimensional $l^1$-regularization optimization, the resulting discretized $L^1$-norm does not have a decoupled form when the standard piecewise linear finite element is employed to discretize the continuous problem. A common approach to overcome this difficulty is employing a nodal quadrature formula to approximately discretize the $L^1$-norm. It is inevitable that this technique will incur an additional error. Different from the traditional approach, a duality-based approach and an accelerated block coordinate descent (ABCD) method is introduced to solve this type of problem via its dual. Based on the discretized dual problem, a new discretized scheme for the $L^1$-norm is presented. Compared new discretized scheme for $L^1$-norm with the nodal quadrature formula, the advantages of our new discretized scheme can be demonstrated in terms of the approximation order. More importantly, finite element error estimates results for the primal problem with the new discretized scheme for the $L^1$-norm are provided, which confirm that this approximation scheme will not change the order of error estimates.

math.OC↗
arXiv2017

An efficient duality-based approach for PDE-constrained sparse optimization

In this paper, elliptic optimal control problems involving the $L^1$-control cost ($L^1$-EOCP) is considered. To numerically discretize $L^1$-EOCP, the standard piecewise linear finite element is employed. However, different from the finite dimensional $l^1$-regularization optimization, the resulting discrete $L^1$-norm does not have a decoupled form. A common approach to overcome this difficulty is employing a nodal quadrature formula to approximately discretize the $L^1$-norm. It is clear that this technique will incur an additional error. To avoid the additional error, solving $L^1$-EOCP via its dual, which can be reformulated as a multi-block unconstrained convex composite minimization problem, is considered. Motivated by the success of the accelerated block coordinate descent (ABCD) method for solving large scale convex minimization problems in finite dimensional space, we consider extending this method to $L^1$-EOCP. Hence, an efficient inexact ABCD method is introduced for solving $L^1$-EOCP. The design of this method combines an inexact 2-block majorized ABCD and the recent advances in the inexact symmetric Gauss-Seidel (sGS) technique for solving a multi-block convex composite quadratic programming whose objective contains a nonsmooth term involving only the first block. The proposed algorithm (called sGS-imABCD) is illustrated at two numerical examples. Numerical results not only confirm the finite element error estimates, but also show that our proposed algorithm is more efficient than (a) the ihADMM (inexact heterogeneous alternating direction method of multipliers), (b) the APG (accelerated proximal gradient) method.

math.OC↗
arXiv2017

Bayesian Model Averaging for the X-Chromosome Inactivation Dilemma in Genetic Association Study

X-chromosome is often excluded from the so called `whole-genome' association studies due to its intrinsic difference between males and females. One particular analytical challenge is the unknown status of X-inactivation, where one of the two X-chromosome variants in females may be randomly selected to be silenced. In the absence of biological evidence in favour of one specific model, we consider a Bayesian model averaging framework that offers a principled way to account for the inherent model uncertainty, providing model averaging-based posterior density intervals and Bayes factors. We examine the inferential properties of the proposed methods via extensive simulation studies, and we apply the methods to a genetic association study of an intestinal disease occurring in about twenty percent of Cystic Fibrosis patients. Compared with the results previously reported assuming the presence of inactivation, we show that the proposed Bayesian methods provide more feature-rich quantities that are useful in practice.

stat.AP↗
arXiv2017

Deep Latent Dirichlet Allocation with Topic-Layer-Adaptive Stochastic Gradient Riemannian MCMC

It is challenging to develop stochastic gradient based scalable inference for deep discrete latent variable models (LVMs), due to the difficulties in not only computing the gradients, but also adapting the step sizes to different latent factors and hidden layers. For the Poisson gamma belief network (PGBN), a recently proposed deep discrete LVM, we derive an alternative representation that is referred to as deep latent Dirichlet allocation (DLDA). Exploiting data augmentation and marginalization techniques, we derive a block-diagonal Fisher information matrix and its inverse for the simplex-constrained global model parameters of DLDA. Exploiting that Fisher information matrix with stochastic gradient MCMC, we present topic-layer-adaptive stochastic gradient Riemannian (TLASGR) MCMC that jointly learns simplex-constrained global parameters across all layers and topics, with topic and layer specific learning rates. State-of-the-art results are demonstrated on big data sets.

stat.ML↗
arXiv2017

A Network Game of Dynamic Traffic

We study a network congestion game of discrete-time dynamic traffic of atomic agents with a single origin-destination pair. Any agent freely makes a dynamic decision at each vertex (e.g., road crossing) and traffic is regulated with given priorities on edges (e.g., road segments). We first constructively prove that there always exists a subgame perfect equilibrium (SPE) in this game. We then study the relationship between this model and a simplified model, in which agents select and fix an origin-destination path simultaneously. We show that the set of Nash equilibrium (NE) flows of the simplified model is a proper subset of the set of SPE flows of our main model. We prove that each NE is also a strong NE and hence weakly Pareto optimal. We establish several other nice properties of NE flows, including global First-In-First-Out. Then for two classes of networks, including series-parallel ones, we show that the queue lengths at equilibrium are bounded at any given instance, which means the price of anarchy of any given game instance is bounded, provided that the inflow size never exceeds the network capacity.

cs.GT↗
arXiv2017

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

We present a class of efficient models called MobileNets for mobile and embedded vision applications. MobileNets are based on a streamlined architecture that uses depth-wise separable convolutions to build light weight deep neural networks. We introduce two simple global hyper-parameters that efficiently trade off between latency and accuracy. These hyper-parameters allow the model builder to choose the right sized model for their application based on the constraints of the problem. We present extensive experiments on resource and accuracy tradeoffs and show strong performance compared to other popular models on ImageNet classification. We then demonstrate the effectiveness of MobileNets across a wide range of applications and use cases including object detection, finegrain classification, face attributes and large scale geo-localization.

cs.CV↗
arXiv2017

Fast Simulation of Hyperplane-Truncated Multivariate Normal Distributions

We introduce a fast and easy-to-implement simulation algorithm for a multivariate normal distribution truncated on the intersection of a set of hyperplanes, and further generalize it to efficiently simulate random variables from a multivariate normal distribution whose covariance (precision) matrix can be decomposed as a positive-definite matrix minus (plus) a low-rank symmetric matrix. Example results illustrate the correctness and efficiency of the proposed simulation algorithms.

stat.CO↗
arXiv2017

Enhanced Piezoelectric Response in Hybrid Lead Halide Perovskite Thin Films via Interfacing with Ferroelectric PbZr$_{0.2}$Ti$_{0.8}$O$_3$

We report a more than 10-fold enhancement of the piezoelectric coefficient $d_{33}$ of polycrystalline CH$_3$NH$_3$PbI$_3$ (MAPbI$_3$) films when interfacing them with ferroelectric PbZr$_{0.2}$Ti$_{0.8}$O$_3$ (PZT). Piezo-response force microscopy (PFM) studies reveal $d_{33}^{MAPbI_3}$ values of ~0.3 pm/V for MAPbI$_3$ deposited on Au and ITO surfaces, with small phase angles fluctuating at length scales smaller than the grain size. In sharp contrast, on samples prepared on epitaxial PZT films, we observe large scale polar domains exhibiting clear, close to 90° PFM phase angles, pointing to polar axes along the film normal. By separating the contributions from the MAPbI$_3$ and PZT layers, we extract a significantly enhanced $d_{33}^{MAPbI_3}$ value of ~4 pm/V, which is attributed to the long-range dipole-dipole interaction facilitated domain nucleation. We also discuss the effect of the interfacial screening layer on the preferred polar direction.

cond-mat.mtrl-sci↗
arXiv2016

Self-dual binary codes from small covers and simple polytopes

We explore the connection between simple polytopes and self-dual binary codes via the theory of small covers. We first show that a small cover $M^n$ over a simple $n$-polytope $P^n$ produces a self-dual code in the sense of Kreck-Puppe if and only if $P^n$ is $n$-colorable and $n$ is odd. Then we show how to describe such a self-dual binary code in terms of the combinatorial information of $P^n$. Moreover, we can define a family of binary codes $\mathfrak{B}_k(P^n)$, $0\leq k\leq n$, from an arbitrary simple $n$-polytope $P^n$. We will give some necessary and sufficient conditions for $\mathfrak{B}_k(P^n)$ to be a self-dual code. A spinoff of our study of such binary codes gives some new ways to judge whether a simple $n$-polytope $P^n$ is $n$-colorable in terms of the associated binary codes $\mathfrak{B}_k(P^n)$. In addition, we prove that the minimum distance of the self-dual binary code obtained from a $3$-colorable simple $3$-polytope is always $4$.

math.AT↗
arXiv2016

Gamma Belief Networks

To infer multilayer deep representations of high-dimensional discrete and nonnegative real vectors, we propose an augmentable gamma belief network (GBN) that factorizes each of its hidden layers into the product of a sparse connection weight matrix and the nonnegative real hidden units of the next layer. The GBN's hidden layers are jointly trained with an upward-downward Gibbs sampler that solves each layer with the same subroutine. The gamma-negative binomial process combined with a layer-wise training strategy allows inferring the width of each layer given a fixed budget on the width of the first layer. Example results illustrate interesting relationships between the width of the first layer and the inferred network structure, and demonstrate that the GBN can add more layers to improve its performance in both unsupervisedly extracting features and predicting heldout data. For exploratory data analysis, we extract trees and subnetworks from the learned deep network to visualize how the very specific factors discovered at the first hidden layer and the increasingly more general factors discovered at deeper hidden layers are related to each other, and we generate synthetic data by propagating random variables through the deep network from the top hidden layer back to the bottom data layer.

stat.ML↗
arXiv2016

Seeing into Darkness: Scotopic Visual Recognition

Images are formed by counting how many photons traveling from a given set of directions hit an image sensor during a given time interval. When photons are few and far in between, the concept of `image' breaks down and it is best to consider directly the flow of photons. Computer vision in this regime, which we call `scotopic', is radically different from the classical image-based paradigm in that visual computations (classification, control, search) have to take place while the stream of photons is captured and decisions may be taken as soon as enough information is available. The scotopic regime is important for biomedical imaging, security, astronomy and many other fields. Here we develop a framework that allows a machine to classify objects with as few photons as possible, while maintaining the error rate below an acceptable threshold. A dynamic and asymptotically optimal speed-accuracy tradeoff is a key feature of this framework. We propose and study an algorithm to optimize the tradeoff of a convolutional network directly from lowlight images and evaluate on simulated images from standard datasets. Surprisingly, scotopic systems can achieve comparable classification performance as traditional vision systems while using less than 0.1% of the photons in a conventional image. In addition, we demonstrate that our algorithms work even when the illuminance of the environment is unknown and varying. Last, we outline a spiking neural network coupled with photon-counting sensors as a power-efficient hardware realization of scotopic algorithms.

cs.CV↗
arXiv2016

The Lifecycle and Cascade of WeChat Social Messaging Groups

Social instant messaging services are emerging as a transformative form with which people connect, communicate with friends in their daily life - they catalyze the formation of social groups, and they bring people stronger sense of community and connection. However, research community still knows little about the formation and evolution of groups in the context of social messaging - their lifecycles, the change in their underlying structures over time, and the diffusion processes by which they develop new members. In this paper, we analyze the daily usage logs from WeChat group messaging platform - the largest standalone messaging communication service in China - with the goal of understanding the processes by which social messaging groups come together, grow new members, and evolve over time. Specifically, we discover a strong dichotomy among groups in terms of their lifecycle, and develop a separability model by taking into account a broad range of group-level features, showing that long-term and short-term groups are inherently distinct. We also found that the lifecycle of messaging groups is largely dependent on their social roles and functions in users' daily social experiences and specific purposes. Given the strong separability between the long-term and short-term groups, we further address the problem concerning the early prediction of successful communities. In addition to modeling the growth and evolution from group-level perspective, we investigate the individual-level attributes of group members and study the diffusion process by which groups gain new members. By considering members' historical engagement behavior as well as the local social network structure that they embedded in, we develop a membership cascade model and demonstrate the effectiveness by achieving AUC of 95.31% in predicting inviter, and an AUC of 98.66% in predicting invitee.

cs.SI↗
arXiv2015

The Poisson Gamma Belief Network

To infer a multilayer representation of high-dimensional count vectors, we propose the Poisson gamma belief network (PGBN) that factorizes each of its layers into the product of a connection weight matrix and the nonnegative real hidden units of the next layer. The PGBN's hidden layers are jointly trained with an upward-downward Gibbs sampler, each iteration of which upward samples Dirichlet distributed connection weight vectors starting from the first layer (bottom data layer), and then downward samples gamma distributed hidden units starting from the top hidden layer. The gamma-negative binomial process combined with a layer-wise training strategy allows the PGBN to infer the width of each layer given a fixed budget on the width of the first layer. The PGBN with a single hidden layer reduces to Poisson factor analysis. Example results on text analysis illustrate interesting relationships between the width of the first layer and the inferred network structure, and demonstrate that the PGBN, whose hidden units are imposed with correlated gamma priors, can add more layers to increase its performance gains over Poisson factor analysis, given the same limit on the width of the first layer.

stat.ML↗
arXiv2015

HiFlash: A History Independent Flash Device

Retention regulations require timely and irrecoverable disposal of data, a challenging task, as data and its side effects are stored and maintained at all layers of a computing system. Those side effects can be used as an oracle to derive the past existence of deleted data. Fortunately, history independence can be utilized to eliminate such history-related oracles. HIFS can provide history independence for file storage over mechanical disk drives. However, HIFS cannot provide history independence when deployed on top of flash devices, as flash memory manages its own internal block placement, which is often inherently history dependent. In this work, we initiate research on history independent flash devices. We design HiFlash, which achieves a strong notion of history independence by defending against an adversary allowed access to the flash at multiple different points in time. In addition, we design a simple, yet history independence friendly wear-leveling mechanism that allows HiFlash to smartly and advantageously trade off a tunable small amount of history leakage for a significant increase in the device's lifetime. Our prototype built in an actual flash device as well as extensive simulations validate the effectiveness of HiFlash.

cs.CR↗
arXiv2015

ELDA: Towards Efficient and Lightweight Detection of Cache Pollution Attacks in NDN

As a promising architectural design for future Internet, named data networking (NDN) relies on in-network caching to efficiently deliver name-based content. However, the in-network caching is vulnerable to cache pollution attacks (CPA), which can reduce cache hits by violating cache locality and significantly degrade the overall performance of NDN. To defend against CPA attacks, the most effective way is to first detect the attacks and then throttle them. Since the CPA attack itself has already imposed a huge burden on victims, to avoid exhausting the remaining resources on the victims for detection purpose, we expect a lightweight detection solution. We thus propose ELDA, an Efficient and Lightweight Detection scheme against cache pollution Attacks, in which we design a Lightweight Flajolet-Martin (LFM) sketch to monitor the interest traffic. Our analysis and simulations demonstrate that, by consuming a few computation and memory resources, ELDA can effectively and efficiently detect CPA attacks.

cs.CR↗
arXiv2015

Structural Properties of an Open Problem in Preemptive Scheduling

Structural properties of optimal preemptive schedules have been studied in a number of recent papers with a primary focus on two structural parameters: the minimum number of preemptions necessary, and a tight lower bound on `shifts', i.e., the sizes of intervals bounded by the times created by preemptions, job starts, or completions. So far only rough bounds for these parameters have been derived for specific problems. This paper sharpens the bounds on these structural parameters for a well-known open problem in the theory of preemptive scheduling: Instances consist of in-trees of $n$ unit-execution-time jobs with release dates, and the objective is to minimize the total completion time on two processors. This is among the current, tantalizing `threshold' problems of scheduling theory: Our literature survey reveals that any significant generalization leads to an NP-hard problem, but that any significant simplification leads to tractable problem. For the above problem, we show that the number of preemptions necessary for optimality need not exceed $2n-1$; that the number must be of order $Ω(\log n)$ for some instances; and that the minimum shift need not be less than $2^{-2n+1}$. These bounds are obtained by combinatorial analysis of optimal schedules rather than by the analysis of polytope corners for linear-program formulations, an approach to be found in earlier papers. The bounds immediately follow from a fundamental structural property called `normality', by which minimal shifts of a job are exponentially decreasing functions. In particular, the first interval between a preempted job's start and its preemption is a multiple of 1/2, the second such interval is a multiple of 1/4, and in general, the $i$-th preemption occurs at a multiple of $2^{-i}$. We expect the new structural properties to play a prominent role in finally settling a vexing, still-open question of complexity.

cs.DM↗