Share on Facebook Tweet on Twitter Share on LinkedIn Share by email
Microsoft Research Asia Fellows

Microsoft Research Asia recognizes the following fellows, who represent the best and the brightest PhD candidates in the Asia-Pacific region.

On this page

2015 fellows 

Wei Bai

Wei Bai

The Hong Kong University of Science and Technology


Supervisor: Kai Chen


Research interests
Data center networking


Long-term research goal
Many data centers have been built around the world to provide various services and applications. In data centers, hundreds of thousands of servers are interconnected by using data center networks. My long-term research goal is to improve the performance of data center networks.


To achieve this goal, I am focusing on designing new transport mechanisms to satisfy the low latency requirements of today’s cloud applications. In the future, I plan to continue working on data center transport on two fronts: designing new solutions for multi-queue multi-service data centers as well as improving network performance for large-scale Remote Direct Memory Access deployments.


Dong DengDong Deng

Tsinghua University


Supervisors: Jianhua Feng, Guoliang Li  

Research interests

Database, data management, data cleaning, and data integration


Long-term research goal

My research focuses on data cleaning, i.e., how to deal with errors and inconsistencies in information systems. As an example, in many applications such as data integration, commercial organizations need to collect data from various sources to conduct analysis and make decisions. The data from these different sources typically contain inconsistencies. For example, the same person’s name may be misspelled or have different formats, such as with or without middle name. Such inconsistencies make it more challenging to link records from different places and answer queries approximately. My long-term research goal is to develop algorithms in order to make query answering and information retrieval efficient in the presence of such inconsistencies and errors.


Li DongLi Dong

Beihang University


Supervisors: Wei Li, Ke Xu

Research interests

Natural language processing, artificial intelligence


Long-term research goal

My research goal is to develop novel and effective techniques for natural language processing, especially question answering. In order to build better natural language interfaces for machines, we need to learn to sufficiently represent text from large volumes of data, and bridge the gap between natural language and structured data. I am interested in leveraging extensive unannotated data in a weak supervision setting, and building neural network-based frameworks to attack the problems of question answering.


Hojoon LeeHojoon Lee

Korea Advanced Institute of Science and Technology


Supervisor: Brent Byunghoon Kang

Research interests

OS security, virtualization


Long-term research goal

My research goal is to enhance the security and reliability of modern systems from the operating system level. I am particularly interested in building secure computing architectures by using trusted execution environments (TEEs). Modern computing systems are under severe threat: software vulnerabilities are constantly discovered and malwares are infecting billions. TEEs provide a safe execution environment in which system integrity monitoring or system transactions can be performed.


I am currently exploring various forms of secure kernel design that involves TEE. My goal is to make fundamental changes to the traditional kernel architectures such that it becomes more resilient against software attacks.


Tae-Hyun OhTae-Hyun Oh

Korea Advanced Institute of Science and Technology


Supervisor: In So Kweon


Research interests

Computer vision and machine learning (especially low-rank and sparse structure)


Long-term research goal

My research interest is to investigate favorable non-convex models for tremendous applications. It has been known that global optimal solutions obtained from convex formulations may not often correspond to physically desirable solutions under many practical conditions. Since there is no limit of convexity while formulating non-convex models, flexible designs can be considered and the working range of algorithms can be broadened.


In this stream, my future research will focus on: 1) Non-convex modeling that has favorable properties. It can be achieved by carefully looking into a problem and data. 2) Developing scalable optimizers corresponding to the newly developed formulations. I would like to make a stepping stone between engineering and theoretical people by providing useful and easy solutions.


Yingwei PanYingwei Pan

University of Science and Technology of China


Supervisor: Houqiang Li 

Research interests

Video understanding, large-scale visual search, click-through data analysis


Long-term research goal

One of the fundamental problems in vision and multimedia is to bridge the semantic gap between visual content and text. In our previous research, we studied the problem from the viewpoint of cross-view (visual-semantic) embedding, that is, integrate content, structure, and/or click data for learning a joint space that connects different natures of vision and text.


Following the research in this direction, I will further explore the complex semantic relationship between long video sequence and natural language by considering three main challenges: 1) learning an effective representation of a long video, 2) video understanding from individual tags to natural sentence, and 3) demonstrating the success of our technologies in a wide range of applications such as video temporal localization and sentence (story) generation.


Hao TangHao Tang

Peking University


Supervisor: Lu Zhang

Research interests

Software analysis and testing


Long-term research goal

Modern programming language features (e.g., callbacks, reflection, file I/O) provide convenient language support for user-defined and platform-specific external behaviors. However, it is difficult to analyze unknown external behaviors with static analyzers. By now, we have made a small step towards analyzing these behaviors. I propose conditional reachability (formulated as tree-adjoining-language reachability) to handle callbacks when performing reachability analysis for software libraries.


In my future work, I plan to improve the efficiency, scalability, precision, and generality of conditional reachability and apply it to real-life mobile and web applications. Also, I will continue to exploit new algorithms and data structures to analyze other types of unknown external behaviors.


Tatsunori TaniaiTatsunori Taniai

The University of Tokyo


Supervisor: Yoichi Sato

Research interests

Mathematical optimization for low- and mid-level computer vision


Long-term research goal

I am interested in applications of mathematical optimization techniques for low- and mid-level computer vision problems, such as 3D shape recovery from images, dense correspondence estimation, and image segmentation. My long-term research goal is to establish a joint framework of image co-segmentation and dense correspondence estimation where common object regions between input images are precisely segmented and aligned as output. Such rich information about object regions and positions is useful in many applications such as fine-grained object detection, scene parsing, and image editing.


Hao WangHao Wang

The Hong Kong University of Science and Technology


Supervisor: Dit-Yan Yeung

Research interests

Machine learning, data mining, Bayesian deep learning, social network analysis, recommender systems


Long-term research goal

There are two primary classes of tasks in AI. The first class is perception tasks like visual object recognition and text understanding. These are the tasks that can be easily completed by normal human beings. The second class is inference and planning. This involves a higher level of intelligence like decision making, data analysis, and logic deduction. In a sense, deep learning with its multiple processing layers is better at the first class of tasks and probabilistic graphical models with its Bayesian nature excel at the second class of tasks. The problem is that in the real world, these two classes often entangle with each other. Our research goal is to combine the power of deep learning and Bayesian models in a single principled framework to get the best of both worlds. We call this framework Bayesian Deep Learning. So far, we have successfully applied Bayesian Deep Learning for applications like recommendation, relational learning, and link prediction. In the future, we would like to extend our work both horizontally, to handle more applications, and vertically, to provide more theoretical guarantees along with more sophisticated models.


Feng-Long XieFeng-Long Xie

Harbin Institute of Technology


Supervisors: Haifeng Li, Frank K. Soong

Research interests

Cross-lingual speech synthesis, voice conversion, speaker adaptation


Long-term research goal

Cross-lingual TTS (text-to-speech) synthesis can be defined as synthesizing speech in the target language (L2) with a specific speaker’s recorded speech in his native language (L1) and to maintain this speaker’s voice characteristics. Currently, I am trying to use Speaker Independent Deep Neural Networks to equalize the speaker difference and Kullback-Leilbler divergence to measure the acoustic units’ difference. In the future, I will apply this framework to many speech applications, e.g., voice conversion, speaker adaptation, speech recognition, language learning, and talking head.


Ting ZhangTing Zhang

University of Science and Technology of China


Supervisors: Yong Wang , Hsiao-Wuen Hon 

Research interests

Computer vision, machine learning


Long-term research goal 

My research interests have revolved around computer vision and machine learning, with a specific focus on fast approximate nearest neighbor (ANN) search. The goal of fast ANN search is to find the closest item of the given query accurately as well the fastest of all of the database vectors. To achieve that goal, our previous work uses composite quantization to represent the database vectors by compact codes. My long-term research goal is to continue exploring algorithms to improve the search accuracy as well as the search efficiency of the approximate nearest neighbors. I believe that many applications—such as similar image search, object retrieval, and so on—can benefit from fast ANN search.


Yongfeng ZhangYongfeng Zhang

Tsinghua University


Supervisor: Shaoping Ma 

Research interests

Personalization theories, computational economics, sentiment analysis


Long-term research goal

I believe that academic research should advance itself not only for commercial benefits, but also for the benefit of society. Economists, philosophers, and sociologists devote their lives for a better human society in our physical world, and we as computer scientists should also push the virtual online society towards a more fair and just world. The research that I am targeting now and in the future falls into the emerging research field of computational economics. I especially care about the new frameworks for online services such as e-commerce, online financing, or freelancing based on the maximization of the joint surplus of both service consumers and providers (i.e., the social surplus) so that online services are matched from providers to consumers in a way that benefits the virtual society for a social good. 


Zhenzhe ZhengZhenzhe Zheng

Shanghai Jiao Tong University


Supervisor: Guihai Chen

Research interests

Algorithmic network economics, wireless network, mobile computing, and cloud computing


Long-term research goal

Generally, I am interested in research problems that lie at the interaction of economic and computer science. My recent research interests mainly focus on resource management in computer networks. With the rapid development of distributed systems, networking systems, and cloud computing technology, modern network resource management needs the cooperation of different network entities. However, the goals of selfish and rationale network entities usually conflict with the overall goal of the whole network system. My research goal is to apply the methodology of game theory to design efficient mechanisms for network resource management in different layers of network systems.

I am also interested in designing efficient and economic-robust market systems, such as online ad auctions and data markets.  

2014 fellows 

Szu-Wen Fan

Szu-Wen Fan

Keio University


Supervisor: Masahiko Inami


Research interests
Substitutional reality (SR), human computer interaction (HCI), cross-modal perception, human's perception of reality


Long-term research goal
I am enthusiastic about finding out how humans perceive reality, based on the multi-modal stimulation we receive with our five senses. As we are well aware, our beliefs about reality are limited to what we perceive with our sensory organs, which could be misdirected from true reality. Therefore, we can immerse users in alternate reality experiences by just stimulating authentic sensory feedbacks using a computer. However, these alternate reality experiences are limited to one reality only. Users are in either actual reality or alternate reality. I suggest that we can extend a human’s perception of reality to be composed of more than one reality simultaneously, by focusing on cross-modal perception and seamless transition between actual and alternate reality. I am looking to blend realities:


  1. From remote reality, for immersive telepresence
  2. From past reality, for immersive memories
  3. From wearable augmented sensory reality, for augmenting human ability

Seunghoon HongSeunghoon Hong

Pohang University of Science and Technology


Supervisor: Bohyung Han


Research interests

Computer vision and machine learning


Long-term research goal

I am interested in developing a visual tracking system that can reliably track the target over “long-term” videos. Although substantial progress has been made in the tracking literature, this problem still remains challenging since most existing trackers are susceptible to model drift and temporary tracking failures. To resolve such issues, my previous research has mainly focused on developing reliable graphical models for tracking. By identifying a suitable order of tracking within a video, more reliable tracking can be achieved robust to intermediate failures.


My future research would focus on long-term visual tracking based on high-level context understanding. Many real-world videos, such as movies and TV shows, contain drastic variations of the target and scene, and tracking based only on low-level visual information would inevitably fail with these challenges. I believe that contexture information, such as scene structure, object categories, and so forth, provides additional information to resolve ambiguities in tracking and to find more meaningful structure of the video for tracking.


Wenping HuWenping Hu

University of Science and Technology of China


Supervisor: Yong Wang, Hsiao-Wuen Hon

Research interests

Computer-aided language learning, speaker recognition, speech recognition, deep learning


Long-term research goal

Computer-Aided Language Learning (CALL) systems, powered by the advancement of speech technology, can bridge the gap between the supply and demand of language teachers and have become ubiquitous learning tools with widely used mobile devices. We have built our own CALL system and transferred it into Microsoft Bing dictionary. This system is able to evaluate language learners’ pronunciation quality, automatic detect the mispronunciation error within the word, and give some constructive feedbacks to language learners, e.g., common error patterns.


Currently, I am using deep learning for pronunciation proficiency assessment and deficiency detection. We trained our golden pronunciation model by deep neural networks, investigated more robust pronunciation algorithms, applied transfer-learning for pronunciation deficiency detection and speeded up the real-time system to make it practical. In the future, we will further extend our algorithms for free-style, spontaneous speech assessment, which will have an assessment of learner's spoken pronunciation proficiency based on his or her free talk recordings, just like TOEFL oral test. Without given the read transcription, it will be more challenging, but more helpful and practical for second language learners.


Yanyan JiangYanyan Jiang

Nanjing University


Supervisor: Jian Lu 

Research interests

Testing and analysis of concurrent programs


Long-term research goal

Software quality assurance is becoming a major challenge for the modern software industry, as software products are getting larger, more complex, and parallel. Software testing practice, being the most prevalent approach for improving software quality, is also encountering emerging challenges. I am particularly interested in testing and analysis techniques to improve the quality of concurrent programs.


Recently, researchers have proposed deterministic replay, symbolic analysis, and predictive analysis. These techniques points out a new trend for testing real-world concurrent software. I am conducting my research towards the following goals:


  1. Record and replay a concurrent program execution that can last days long within affordable resource consumption
  2. Design sound and semantic-aware causal execution model for concurrent programs that can be checked in polynomial time
  3. Using symbolic analysis and constraint solving technique to systematically explore interleaving space of concurrent programs

Hiroshi KajinoHiroshi Kajino

The University of Tokyo


Supervisor: Kenji Yamanishi, Hisashi Kashima


Research interests

Human computation, crowdsourcing, machine learning, data mining


Long-term research goal

My research goal is to establish a secure and reliable infrastructure for human computation. Human computation is an emerging computer science technique in which machines utilize humans as computational resources to tackle computationally hard problems. Crowdsourcing is often used to allow machines to access human resources via the web. However, it has been often pointed out that crowdsourcing brings about a reliability issue that a result of computation is sometimes unreliable.


My research focuses not only on the reliability issue, but also on privacy issues that arise, for example, when a task is submitted to crowdsourcing and when workers in crowdsourcing return computation results. By addressing the privacy issues, the applicability of human computation will become wider than ever.


Yuan LinYuan Lin

Fudan University


Supervisor: Zhongzhi Zhang 

Research interests

Random walks in complex network


Long-term research goal

Random walks are a fascinating research topic attracting intensive attention within the scientific community, not only for their intrinsic mathematical beauty but also for their important implications to various scientific issues, such as target searching, node ranking, computer vision, to name just a few. My research goal will focus on the following two tasks.


  1. Discovering the behavior of biased random walks in complex networks. Although unbiased random walks have been widely investigated, many dynamics should be described by biased random walks, which is still an outstanding issue. My work will both numerically and analytically research the influence of different biases on the crucial random walk indicators.
  2. Applying new random walk mechanisms in practical problems. Based on the obtained exhibitions of various biased random walks, my research will explore the applications of such bias in virus spreading, link prediction, community detection, and so on.

Yutao LiuYutao Liu

Shanghai Jiao Tong University


Supervisor: Haibo Chen 

Research interests

Cloud and mobile security


Long-term research goal 

My primary research goal is to improve server and mobile security in system perspective. My current  research projects investigate two areas of focus:


  1. Leveraging potential architecture support to improve the dependability and security of current systems, specifically, using hardware features in different architectures to protect data privacy, enforce control flow integrity, actively monitor the whole system, timely detect potential attack, and so forth.
  2. Building and optimizing reliable virtualization environment to benefit system security. Virtualization has been a tremendous success, especially in server consolidation and security enhancement. However, there are still some problems when considering security aspects, and my goal is to ameliorate virtualization architecture to adapt to increasing security requirements.

Canyi LuCanyi Lu

National University of Singapore


Supervisor: Shuicheng Yan, Zhouchen Lin 

Research interests

Compressive sensing of block diagonal affinity matrix, convex optimization


Long-term research goal

I am interested in developing algorithms for important problems with support in theory. In particular, I am aiming to establish two or three representative works in my PhD career. They include:


  1. Compressive sensing of block diagonal affinity matrix
    Affinity matrix is almost everywhere. The ideal affinity matrix should be block diagonal, i.e., data points from different subspaces or subjects are not similar at all. There are many known methods, such as sparse representation and low-rank representation, which lead to the block diagonal solution under certain conditions. We have established the Enforced Block Diagonal (EBD) conditions to unify many previous methods. We are aiming to find a convex model for block diagonal affinity matrix pursuit and to establish the theory for the exact recovery.
  2. Alternating Direction Method of Multiplier (ADMM)
    ADMM is the most important solver for convex optimization with linear constraint. It regains much attention in recent years due to the success of compressive sensing. There are many ADMM variants based on different structures of the objective function, but their convergence proofs are given case by case. We are aiming to give a general framework that unifies many known ADMMs with a unified proof. Also, there are still many open problems with ADMM. I aim to give a comprehensive study in the future.

I am also interested in applying our algorithms and theory for computer vision and pattern recognition.


Yongtae ParkYongtae Park

Korea University


Supervisor: Hyogon Kim 

Research interests

Software defined radio for mobile devices, vehicular networking


Long-term research goal 

I am working on Software Defined Radio (SDR) for mobile devices. Recently, I demonstrated the feasibility of ZigBee communication on smartphones and am working to show the possibility of supporting various protocols without using dedicated chips. Providing these protocols as an app will catalyze the development and maintenance of protocols on future mobile devices. Also, it will enable people to develop the personalized and secure protocols for themselves.


To make this happen, I am working on various issues such as optimization of SDR processing, fairness of antenna sharing, and so forth. Ultimately, my long-term research goal is to make mobile devices speak thousands of protocols. I hope my research plays an important role to facilitate “Internet of Things” by helping to connect humans with ambient smart devices.


Shaoqing RenShaoqing Ren

University of Science and Technology of China


Supervisor: Bin Li, Jian Sun 

Research interests

Computer vision, detection, and localization for face and generic objects


Long-term research goal 

My research interest lies in computer vision, especially detection and localization. Detection and localization as the fundamental problems in computer vision have been studied widely and are closing to practical applications. My previous research focuses on highly efficient detection and localization (alignment) on human face and generic objects, which makes these computer vision technologies cheaper on computation cost and closer to everybody’s daily life.


I envision more and more applications will benefit from the breakthrough of image understanding, especially apps on mobile devices. My future research will continue to explore accurate and highly efficient detection/localization systems and make them more practical for real applications.


Yingxia ShaoYingxia Shao

Peking University


Supervisor: Bin Cui 

Research interests

Large-scale graph analysis, parallel computing framework, scalable data processing


Long-term research goal

With the rapid growth of social networks, e-commence, and sensor networks, large-scale graphs become easily available in public. These graphs represent the complex relationships in the real world and contain a lot of valuable knowledge. Therefore, mining and analyzing these large-scale graphs is an important approach to discover the knowledge. The long-term goal of my research is to scale graph analysis to the scales of real-world graphs and develop a scalable and efficient large graph analysis toolkit.


To achieve this goal, we design new graph algorithms according to several principles. First, the new algorithms will exploit the new characteristics of graphs, such as sparse, dynamic, and power-law distribution. Second, the workload patterns of graph problems will be unraveled as well. Lastly, for intricate graph metrics (e.g., SimRank, closeness, betweenness, and so forth), we will design new approximation algorithms. The toolkit will be carefully designed by abstracting primitive operations among above graph analysis task.


Shuai YiShuai Yi

The Chinese University of Hong Kong


Supervisor: Xiaogang Wang 

Research interests

Computer vision and crowd video surveillance


Long-term research goal

I am now focusing on crowd video analysis in computer vision. Nowadays, with steady population growth and worldwide urbanization, crowded situations are more common. Crowd management and traffic control are common problems in public areas with a high population density. Millions of surveillance cameras are capturing these areas day and night. However, it is still very challenging to analyze these videos accurately and efficiently.


A pedestrian’s decision-making is very complex and many factors may have great influence on one’s walking behavior. For example, the structure of the scene, the inner belief of the pedestrian, and the interaction with other moving pedestrians and stationary groups will all change pedestrians’ walking paths and make the problem difficult to solve.


I am now focusing my analysis on human behaviors in crowds and trying to find the relationship between a pedestrian’s walking behavior and corresponding influence factors. In the future, I hope a joint pedestrian decision making model can be built which will aid in behavior analysis and abnormal detection.


2013 fellows

Chen CaoChen Cao

Zhejiang University

Supervisor: Kun Zhou

Research interests
Computer graphics, real-time facial animation

Long-term research goal
My research goal is to propose a real-time facial animation system for average users. Several issues need to be addressed to implement such a system. Primarily, the usage of special equipment such as markers or camera arrays should be avoided. Alternatively, an ordinary web camera should be the only capture device. Our system expectations are to be used in a wide range of environments, including both indoor and outdoor environments exposed to direct sunlight. The system should be capable of handling the changes in lighting and background. It is also vital to ensure maximum high performance system capabilities with the system.

I will continue to focus on researching real-time facial animation in the future. There are many problems that need to be addressed and fixed with Real-time Facial Animation. Our current system requires a large amount of training data during the preprocessing step. It is ideal to find a solution to reduce the training data requirements, because the user needs to perform 60 different facial expressions and head poses.

A key goal is to implement our real-time facial animation system on mobile devices. This, however, is not straightforward. Due to the limited computational power on mobile devices, the three-dimensional shape regression is time consuming. The limitations in our current system to handle dramatic lighting changes must be addressed since mobile devices are frequently used in various environments.

Dong ChenDong Chen

University of Science and Technology of China

Supervisor: Bin Li, Jian Sun

Research interests
Computer vision, face recognition, face detection

Long-term research goal
Face recognition has been widely studied for more than three decades. Facial recognition of an image is divided into four steps: face detection, face alignment, feature extraction, and feature learning. Each element will have an impact on the overall performance. My long-term research goal is to improve face recognition accuracy during the four-step process. The following implementations will highlight my long-term research goal.

  1. A more efficient structure for face detection and the face alignment pipeline will be designed. Face recognition’s speed and accuracy, and the Landmark’s Localization Algorithm will be improved.
  2. An illumination invariant feature combined with data-driven based approaches to increase face recognition accuracy under uncontrolled extreme lighting conditions will be designed.
  3. I will extend my feature and learning method to more face-related tasks such as attribute recognition, as well as exploiting these technologies to further promote the recognition rate.

Luwei ChengLuwei Cheng

The University of Hong Kong

Supervisor: Cho-Li Wang, Francis Chi Moon Lau

Research interests
Virtualization technologies for cloud computing

Long-term research goal
Virtualization is largely adopted in datacenters to support on-demand cloud computing. Current virtualization technologies are primarily invented to host unmodified or slightly modified operating systems using virtual machines (VMs). However, the existence of the hypervisor, an additional layer between the hardware and the guest operating system has significantly changed the assumptions of many protocols, which are originally designed for physical running environments. In particular, VM scheduling delays due to CPU sharing can negatively affect many kernel services such as the process scheduler, the disk scheduler and network protocols. Consequently, my research efforts go towards deeply integrating the hypervisor and the guest OS to form a more lightweight container for upper-layer applications.

Quan FangQuan Fang

Chinese Academy of Sciences

Supervisor: Changsheng Xu

Research interests
Georeferenced social media research and application

Long-term research goal
research interest lies in georeferenced social media research and application. I focus on geo-social computing in particular, which aims to sense and mine the massive georeferenced social media data to understand the user and geo-location for better serving.

I am working towards the following strategies.

Georeferenced media mining and modeling aims to sense and mine the massive geo-tagged data to understand geo-location, especially from the human-sensed perspective. User modeling in social media is to understand users by exploiting rich online user-generated content. I will combine and model the user, geo-location, and content in a unified framework.

I am dedicated to developing efficient and effective techniques for harvesting knowledge about users and geo-locations from georeferenced social media data. The derived knowledge can benefit various user-centric applications.

Ping LuoPing Luo

The Chinese University of Hong Kong

Supervisor: Xiaoou Tang

Research interests
Computer vision, computer graphics, and machine learning

Long-term research goal
I am interested in the areas of computer vision, computer graphics, and machine learning. My particular focus is deep learning and its applications with vision and graphics problems, including face and human detection, parsing (poses, clothing), attribute analysis, and identification. I envision that information technology in the future can help automatically find, identify, and connect people all over the world from various media—such as photographs, social networking, and images on the Internet—making everybody’s life easier. My future research heads toward this goal both theoretically and practically.

Image representation is a fundamental problem in computer vision. Although this representation can be directly learned from data by deep learning, how to learn or fuse the representations from different domains is still a great challenge. I will first devise a cross-domain multi-task method for image modeling so that it can fuse the data from different sources, while serving as the building block for many high-level tasks; such as detection, recognition, and segmentation.

High computational cost is another challenge when working with large amounts of data. My second task will focus on building a large-scale distributed deep-learning framework, which can effectively deal with tens of millions of image data.

Masayasu OgataMasayasu Ogata

Keio University

Supervisor: Michita Imai

Research interests
Human-computer interaction, natural user interface

Long-term research goal
I am interested in the area of HCI, especially natural user interface and the embodied interaction with sensing technique design and development, as well as the wearable device and interactive system. Our experiences with the living environment are getting complicated with user motivation and desired functionality. To ease and illuminate interfaces in daily life, I will propose a vision and an idea to increase interface transparency that makes the user unaware of the computer’s existence when they access the everyday environment. Sensing techniques and designing appropriate feedback to the user are required for such an achievement. My experience and skill will reinforce the long-term goal that integrates the human and the system in a daily environment in the near future.

Jianping ShiJianping Shi

The Chinese University of Hong Kong

Supervisor: Jiaya Jia

Research interests
Computer vision, machine learning

Long-term research goal
I have an extensive interest in computer vision and machine learning. My previous research focuses mainly on the model’s point of view, including Sparse and Low-rank Matrix Factorization. These models have been widely used in many applications. To be proficient in modeling is not enough. I believe building appropriate models to fit the vision data is much more important. I am currently working on several projects, including Image Blur Analysis, Human Detection, and RGBD Video Processing. Such projects necessitate a deep understanding of the image data. For example, the Blur Analysis is based on several natural image statistics. Our Human Co-detection System explores the intrinsic similarity for the same person across photos. The RGBD Visual Tracking utilizes the depth distribution of the tracking object. All these projects stem from vision data understanding. My future research will continue to explore computer vision’s interesting aspects, while also advancing the computer to understand our colorful world.

Shusen WangShusen Wang

Zhejiang University

Supervisor: Zhihua Zhang

Research interests
Machine learning, matrix analysis, convex optimization

Long-term research goal
My research focus is on enabling large-scale machine learning using randomized approximations. As we know, many classical machine learning methods suffer from high time and space complexities and are thus prohibitive in big-data applications. Some randomized approximation approaches—for example, random projection, random selection, and the Nystrom method—reduce the time and space complexities of many machine-learning methods from quadric or cubic to near linear by sacrificing a little accuracy. I am working on improving the existing approximation methods, devising new approximation methods, and providing theoretical guarantees for the methods. My goal is to make the randomized approximation approaches more accurate and more efficiently computed.

Huanjing YueHuanjing Yue

Tianjin University

Supervisors: Feng Wu, Jingyu Yang

Research interests
Cloud-based image coding and processing

Long-term research goal
My long-term research goal is to advance the state-of-the-art technologies in signal processing and computer vision through cloud data. The number of images shared on the Internet is dramatically increasing due to the dramatic increase in social websites. This provides a live and huge image base, which has posed demands for the development and implementation of content-based image retrieval. In addition, it brings new opportunities for many well-known, ill-posed image processing and computer vision problems, such as image composition, completion, and colorization. I mainly focus on cloud-based image coding, super-resolution, and denoising. Different from the traditional methods, our work proposes utilizing the external correlated images to improve coding efficiency and image restoration quality. Our work is an early step toward a new era of cloud-based image coding and processing.

Jun ZhangJun Zhang

Nanyang Technological University

Supervisor: Xiaokui Xiao

Research interests
Database management and data privacy

Long-term research goal
There is a growing emphasis in the world at large on utilizing and disseminating aggregate statistics from demographic data, health records, Internet activity, and other sources. This data, however, can seldom be accessed for public studies due to concerns over individual privacy. Existing methods for privacy preservation still offers inadequate assurance, because they rely on restrictive assumptions on how a malicious user may attack the data, whereas those assumptions can be easily violated in practice.

Motivated by the practical importance of privacy protection and the deficiencies of the existing methods, my research work aims to conduct a comprehensive study on mathematically rigorous privacy protection models. Specifically, I am interested in:

  1. Generic and fundamental algorithms under Differential Privacy, a state-of-the-art privacy model that is rigorously formulated based on statistics, offering an extremely strong privacy guarantee;
  2. Novel class of privacy definitions with better trade-off between privacy and utility.

2012 fellows

Yang CaoYang Cao

Beihang University

Supervisor: Jinpeng Huai

Research interests
Database theory and systems, graph pattern matching, social searching

Long-term research goal

  1. Graph pattern matching revised for social searching: Graph pattern matching is fundamental to social analysis. Traditional techniques like subgraph isomor-phism and graph simulation either impose too strong a topological constraint on graphs to retrieve meaningful matches and incur high computational complexity (for example, subgraph isomorphism is NP-complete), or are too loose to find correct matches. Moreover, data graphs for modeling social networks are usually evolving over time. Existing models and techniques often fall short of handling this without incurring high computational complexity. To this end, I will propose new graph pattern matching models (queries) as well as techniques (algorithms) for evaluating them. The model should be well-designed such that it could strike a balance between the express power (semantics) and the computational complexity for evaluating them, and should be able to handle the evolvement of data graphs as well.
  2. Keywords search on graph data: Keywords searching has been a longstanding issue for search engines. Traditionally, keyword searching is often conducted on structured and semistructured data, as well as documents. Existing models and techniques for keywords searching cannot be naively adopted for social search engines. Recent research about keywords searching on graph data mainly focuses on developing techniques for retrieving matched results on graph data. My research for this aspect will instead turn to proposing new keywords searching models specific for graph data, to take the advantages of graph pattern queries to provide more accurate semantics for emerging applications.
  3. New computational model and complexity classes: In the context of large-scale data, traditional computational classes are too generic for identifying problems that are solvable practically, and moreover, they cannot take properties of data graphs (social networks) into consideration. To this end, I will propose a computational model and a hierarchy of complexity classes, to provide a dichotomy between those queries that are feasible on large-scale data graphs and those that are not, and thus properly classify social searching problems, with regard to the express power and evaluating complexity of the queries.

Menglei ChaiMenglei Chai

Zhejiang University

Supervisor: Kun Zhou

Research interests
Computer Graphics, Computer Vision, Image Processing, and so forth

Long-term research goal
My long-term research goal primarily targets portrait image manipulation, especially for recovering semantic information from common images and providing a series of techniques which enables useful applications for average users. As we know, images and videos featuring humans as their subjects are of great interest for both industrial experts and average end users, and have motivated various kinds of research in computer graphics during the past decades. But few of them involve 3-D information to resolve the occlusions and ambiguities issues, making these low-level image processing methods of limited application for specific tasks of portrait manipulation. Our research work will focus on portrait image manipulation, especially for components such as face, hair, body shape, clothing, and so forth. The key thought is to extract 3-D information by utilizing specific prior knowledge achieved with machine-learning techniques, and apply them to drive a broad range of user-friendly applications which were previously challenging.

Jun KatoJun Kato

The University of Tokyo

Supervisor: Takeo Igarashi

Research interests

Human-computer interaction and programming language (user interface for programmers)

Long-term research goal
I am interested in the broad area of human-computer interaction, but have been especially focused on interactions between humans and the real world through computers (physical computing and robot applications), their input modalities (natural user interfaces), and their development methods (prototyping toolkits and integrated development environments). In my future vision, everyone can be a weekend carpenter who makes his/her everyday life easy and comfortable with help of information technology—in other words, end-user programming of the real world in the real world. Currently, I am planning to take two steps toward this goal.

  • First, I will propose development environments that improve the programmer’s experience (Programmer’s eXperience, PX) on “real-world programming” that involves interactions with the real world, such as robot control and vision-based object recognition.
  • Second, I will propose end-user development environments for the same “real-world programming” that make creation of customized applications feasible to end users through specialized user interfaces.

Given my past experience on PX research, I expect that further investigation on the professional user interfaces will serve as the fundamentals for designing new user interfaces for the end users.

Yin-Hsi KuoYin-Hsi Kuo

National Taiwan University

Supervisor: Winston H. Hsu

Research interests
Multimedia content analysis, image retrieval, and mobile visual search

Long-term research goal
Nowadays, the rapid computational power of mobile devices brings the emerging need for mobile visual search. Different from the traditional content-based retrieval system, the mobile devices have the ability to process the query (for example, feature extraction) before the transmission. In recent years, the hash-based method has become promising for approximate nearest neighbor (ANN) search because of its ability to deal with high-dimensional features and large-scale databases in an efficient way. An efficient and effective method on the mobile devices is extremely crucial; hence, the hash-based approach becomes a possible direction in reality. In the future, we attempt to integrate the hash-based approach with contextual information to achieve efficient and scalable mobile visual search.

Hongjin LiangHongjin Liang

University of Science and Technology of China

Supervisor: Xinyu Feng

Research interests
Program verification

Long-term research goal
The computer industry has shown explosive growth in the past years. This trend is not likely to slow down. However, the lack of reliable and secure software is becoming a bottleneck for such growth. In particular, programming on multiprocessors is extremely error-prone, but very difficult to debug. Thus, in my long-term research career, I would like to develop new and easy-to-use technologies and tools that could improve the reliability, safety, and security of software on multiprocessors. I am particularly interested in:

  1. Designing program logics for verifying the correctness (including safety and liveness properties) of concurrent programs
  2. Developing automatic or semi-automatic tools to support concurrent program verification
  3. Applying my technologies and tools to verify today’s concurrent libraries (such as java.concurrent.util), runtime support in concurrent systems (such as concurrent garbage collectors), and other concurrent software used in practice

Yongxin TongYongxin Tong

Hong Kong University of Science and Technology

Supervisor: Lei Chen

Research interests
Database and data mining, machine learning

Long-term research goal
My research interests mainly focus on uncertain data management and mining. With the emergence of many real applications—such as sensor network monitoring, moving object search, protein-protein interaction (PPI) network analysis, and so forth—managing and mining uncertain data has attracted much attention in the database and data mining communities recently. Although current researches have solved some fundamental operations over uncertain data—for example, join, ranking, mining frequent item sets, clustering, and so forth—there are a few works to explore the hidden correlation in uncertain data due to the intricate probabilistic structure and high computational complexity. Therefore, my research aims to address these challenges via incorporating both theoretical and practical viewpoints.

  • On the theoretical side, I will first design a novel model to capture intrinsic correlated properties in uncertain data. Additionally, I will propose a series of efficient and effective algorithms in order to adapt complex structural data, such as uncertain data streams and uncertain graphs. This is particularly important in the age of big data as well.
  • On the practical side, I hope to extend our theoretical model and algorithms to real application scenarios, in other words, I try to develop a new crowdsourcing platform based on mobile computing and utilize our probabilistic model to handle the uncertainty control in this system, which is one of most urgent problems in the crowdsourcing researches.

In summary, my research goal is to better discover and manage the hidden correlation and rules over massive uncertain data effectively.

Xinggang WangXinggang Wang

Huazhong University of Science and Technology

Supervisor: Wenyu Liu

Research interests
Computer vision and machine learning

Long-term research goal
My research interests are computer vision and machine learning, especially the problem of object detection. Object detection concerning detecting instances of semantic objects of a certain class in digital images and videos, is a fundamental problem in computer vision, and has wide applications in people’s lives, for example, face recognition, image search, automatic driving, and so forth. Through years of research, the problem of frontal face detection has been well studied, and this technology has been integrated into products. However, detecting most of the other objects in the real world is a very difficult problem due to the great variations of object appearances, such as persons and cars. My previous research mainly focuses on part-based object detection, shape-based object detection, and semi-supervised learning for object detection. It brings together shape feature design, object model design, machine learning, and optimization.

Based on my previous research, my long-term research goal is to build an object detection system in a very weakly supervised way, which can achieve the state-of-the-art performance and be used in some real-life applications. I will put more effort on:

  1. Studying discriminative object part detector; for example, people can easily recognize a leopard when looking at dapple of leopard. Part detector is robust to occlusion. Ensemble of part detector can reach good object detector.
  2. Using context information for object detection which combines scene understanding and object recognition.
  3. Combining modeling based methods with data-driven based methods.

Yizhong ZhangYizhong Zhang

Zhejiang University

Supervisor: Kun Zhou

Research interests
Computer graphics, physically based simulation

Long-term research goal
My research interest is in physically based simulation, especially simulation of complex natural phenomena for graphics. Physically based simulation is a powerful tool in both science and engineering. It can help us to predict unknown phenomena that may prevent the success of an experiment or the improvement of a product's design before manufacturing. In graphics, we need simulations to make animations of water, fire, cloth, and many other phenomena that contain high frequency details. We want the simulation to be fast while being reliabile so that it will be easier for artists to control the animation. My research goal is to improve the performance of simulation, including propose reduced physical model, method of efficient and robust handling of geometry, and parallel computing. By using these methods, we can get high quality simulation results more quickly.

Xin ZhaoXin Zhao

Peking University

Supervisor: Xiaoming Li

Research interests
Social media analysis, web text mining, machine learning

Long-term research goal
My general research interests are within the area of online social networks(for example,Twitter and Facebook) analysis. Online social networks—in particular, their rapid growth and development—are continuously attracting users all around the world, which significantly changes the way that people live. Although various text mining methods have been shown effective to deal with traditional document collections, for example, scientific publications, very few of them are tested to achieve very robust and sound performance on real data sets of online social networks due to the fact we still do not fully understand the nature of online social networks, such as their underlying information structure, user behaviors, and connecting patterns.

To overcome these difficulties, the goal of my research is to develop both principled methodologies and innovative applications for automatically analyzing and discovering knowledge from online social networks. Specifically, I will mainly focus on the aspect of content analysis of online social networks in the long term. More challenging than traditional content analysis, we have to first understand the underlying information patterns and uncover the generative process of such information before we can construct effective models. Previous research experiences will be helpful for me to identify and solve real-world problems that are valuable to common users in this direction. During the problem-solving process, I will try to construct formal methods with clear and intuitive motivations, borrow the ideas and techniques from multiple disciplines, and evaluate research results with large-scale real data.

Chong Zu

Tsinghua University

Supervisor: Chi-Chih Yao

Research interests

Quantum information processing

Long-term research goal
Quantum information and computation is an interdisciplinary subject of computer science and quantum physics. During recent decades, there have been a lot of important theoretical results in this field, including unconditional secure cryptography, true random number generation, exponential speedup algorithm for factoring, and so on. However, realizing quantum information processing and quantum computation is still not an easy task. In my research, I am mainly focusing on two different physical systems, photonic qubit and NV center in diamond, both of which are promising candidates for future quantum information processing and computation.

  1. Photonic qubit: In my recent research, we are working on a project to theoretically propose and experimentally build a brand-new, reliable, and practical quantum random number generator. By saying reliable, it means that we can always use some methods to certify that the random number we are producing is indeed from quantum power, rather than some classical device or lack knowledge of the device, and the randomness (characterized by min-entropy) can be bounded; while saying practical means that our random number generator can be efficient and easy to be realized. In the future, we plan to generate quantum state with more qubits, with which we can demonstrate interesting quantum information protocols, as well as realize large-scale quantum information processing.
  2. NV center in diamond: Recently, we have started to build NV center system. We want to use it to realize room-temperature quantum memory, as well as solid state quantum repeater for long-range quantum information. Our long-term goal is to build a hybrid quantum computation and information network with different quantum systems including trapped ion, matter qubit, and photonic qubit, which can be treated as a prototype for a future genuine quantum network.

2011 fellows

  • Dongzhe Ma, Tsinghua University
  • Peiran Ren, Tsinghua University
  • Xiaohui Bei, Tsinghua University
  • Xun Cao, Tsinghua University
  • Quan Wang, Peking University
  • Xiang Song, Fudan University
  • Chenglin Li, Shanghai Jiao Tong University
  • Guo Li, Zhejiang University
  • Shengjun Huang, Nanjing University
  • Jing Yuan, University of Science and Technology of China
  • Cuiling Lan, Xidian University
  • Guangmin Wang, Xi’an Jiao Tong University
  • Yang Cao, University of Science and Technology of Huazhong
  • Qiang Hao, Tianjin University
  • Zhen Liao, Nankai University
  • Lei Cui, Harbin Institute Technology University
  • Xiaoshuai Sun, Harbin Institute Technology University
  • Bolei Zhou, The Chinese University of Hong Kong
  • Lin Ma, The Chinese University of Hong Kong
  • Nobuyuki Umetani, The University of Tokyo
  • Yefeng Liu, Waseda University
  • Hyunson Seo, Yonsei University
  • Yohan Chon, Yonsei University
  • Jaesik Park, Korea Advanced Institute of Science and Technology
  • Sangwon Seo, Korea Advanced Institute of Science and Technology
  • Kazi Rubaiat Habib, National University of Singapore
  • Gang Yu, Nanyang Technological University
  • Nikolay Gravin, Nanyang Technological University
  • Shuo-Hung Chen, National Tsing Hua University

2010 fellows

  • Haris Javaid, The University of New South Wales
  • Yue Deng, Tsinghua University
  • Chongyang Ma, Tsinghua University
  • Xiong Li, Shanghai Jiao Tong University
  • Bo Geng, Peking University
  • Shiliang Zhang, Institute of Computing Technology, The Chinese Academy of Science
  • Xiulian Peng, University of Science and Technology of China
  • Xiao Zhang, Tsinghua University
  • Xinying Song, Harbin Institute Technology University
  • Jinbao Wang, Harbin Institute Technology University
  • Wei Wu, Peking University
  • Linghe Kong, Shanghai Jiao Tong University
  • Liang Wang, Nanjing University
  • Ping Chen, Nanjing University
  • Weiwei Wu, University of Science and Technology of China
  • Xiaojun Qian, The Chinese University of Hong Kong
  • Dongxiao Yu, The University of Hong Kong
  • Lu Wang, The Hong Kong University of Science & Technology
  • Seokhwan Kim, University of Tsukuba
  • Adiyan Mujibiya, The University of Tokyo
  • Seungjin Lee, Korea Advanced Institute of Science and Technology
  • Tae-Joon Kim, Korea Advanced Institute of Science and Technology
  • Gae-won You, Pohang University of Science and Technology
  • Sungjin Lee, Seoul National University
  • Shenghua Gao, Nanyang Technological University
  • Yi-Ling Hsieh, National Cheng Kung University
  • Cheng-Te Li, National Taiwan University
  • Tsung-Te Lai, National Taiwan University

2009 fellows

  • Novi Quadrianto, The Australian National University
  • Ashnil Kumar, The University of Sydney
  • William Voorsluys, The University of Melbourne
  • Shenghua Liu, Tsinghua University
  • Hao Wen, Tsinghua University
  • Zhiwei Xiong, University of Science and Technology of China
  • Dong Liu, Harbin Institute of Technology
  • Bo Yu, Harbin Institute of Technology
  • Litian Tao, Beihang University
  • Yufeng Li, Nanjing University
  • Wei Wang, Nanjing University
  • Huanhuan Cao, University of Science and Technology of China
  • Xiaoyin Wang, Peking University
  • Jun Lang, Beijing Institute of Technology
  • Derek Hao Hu, The Hong Kong University of Science & Technology
  • Kaiming He, The Chinese University of Hong Kong
  • Tasuku Oonishi. Tokyo Institute of Technology
  • Yoshida Yuichi, Kyoto University
  • Jun Hatori, The University of Tokyo
  • Jongwuk Lee, Pohang University of Science and Technology
  • Jung-Tae Lee, Korea University
  • Bingjun Zhang, National University of Singapore
  • Lixin Duan, Nanyang Technological University
  • Yu-Chen Sun, National Chiao Tung University
  • Kai-yin Chen, National Taiwan University

2008 fellows

  • Feng Zhang, Beijing Institute of Technology
  • Xiangyi Meng, Beijing Institute of Technology
  • Tong Wu, Beijing University of Post and Telecommunication
  • Li Xu, Chinese University of Hong Kong
  • Jun Lang, Harbin Institute of Technology
  • Qingqing Zhang, Institute of Acoustics, Chinese Academy of Sciences
  • Yanyan Lan, Institute of Applied Mathematics, Academy of Mathematics and Systems Science, Chinese Academy of Sciences
  • Xiaoqin Zhang, Institute of Automation, Chinese Academy of Sciences
  • Xiubo Geng, Institute of Computing Technology, Chinese Academy of Sciences
  • Souneil Park, Korea Advanced Institute of Science and Technology
  • Ki-woong Park, Korea Advanced Institute of Science and Technology
  • Dafan Dong, Nankai University
  • Yi Huang, Nanyang Technological University
  • Yu-Lin Wang, National Cheng Kung University
  • Yi-Hsuan Yang, National Taiwan University
  • Yantao Zheng, National University of Singapore
  • Lijiang Chen, Peking University
  • Sunghyun Cho, Pohang University of Science and Technology
  • YongDeok Kim, Pohang University of Science and Technology
  • Dikan Xing, Shanghai Jiao Tong University
  • Jingjing Fu, The Hong Kong University of Science and Technology
  • Jun Hong, The University of Hong Kong
  • Florian Mueller, The University of Melbourne
  • Tomoaki Higo, The University of Tokyo
  • Pinyan Lu, Tsinghua University
  • Jialin Zhang, Tsinghua University
  • Qiming Hou, Tsinghua University
  • Kun Xu, Tsinghua University
  • SYifei Don, University of New South Wales
  • Hao Xu, University of Science and Technology of China
  • Yuan Liu, University of Science and Technology of China
  • Myung - Suk Song, Yonsei University