3D medical image classification with deep learning methods is challenging due to two problems, the difficulty of designing the neural architecture (NA) and the lack of large-scale 3D medical datasets for pretraining. Neural architecture... more
3D medical image classification with deep learning methods is challenging due to two problems, the difficulty of designing the neural architecture (NA) and the lack of large-scale 3D medical datasets for pretraining. Neural architecture search (NAS) has become a popular method to design NA automatically. Inspired by NAS, recent studies have proposed automatic data augmentation (ADA) to search the data augmentation policy (DAP) to increase the size of 3D medical datasets. However, NAS or ADA can only solve one of the above two problems. It is difficult to search both DAP and NA for 3D medical images because the computational cost of 3D data augmentation operations is high, and the joint search space is too large to find the optimal solution economically. To this end, we propose an efficient end-to-end framework for the joint search of DAP and NA applicable to 3D medical images, namely \textit{MedPipe}. Specifically, 1) we modify all 3D data augmentation operations to be differentiabl...
Research Interests:
Research Interests:
In today's mobile devices, the battery reservoir remains severely limited in capacity, making power consumption a key concern in the design and implementation of mobile applications. In this paper, we closely examine one widely... more
In today's mobile devices, the battery reservoir remains severely limited in capacity, making power consumption a key concern in the design and implementation of mobile applications. In this paper, we closely examine one widely adopted approach to improve the energy efficiency of mobile applications-adaptively offloading the computation to the remote cloud. In particular, we measure the power consumption of computation offloading for two representative real-world mobile cloud applications under various wireless network conditions and identify the unique features of data transmission for computation offloading. We then formulate the power-aware scheduling problem for computation offloading and present a scheduling algorithm that makes adaptive offloading decisions according to the dynamic network conditions. Simulation results show that our proposed method can achieve better battery performance, which also reveal that computation-intensive and delay-tolerant tasks are more likely...
Research Interests:
Research Interests:
Research Interests:
Research Interests:
Research Interests:
In cognitive radio networks, rendezvous is a fundamental operation by which cognitive users establish communication links. Most of existing works consider rendezvous of a pair of users. When multiple pairs of users are doing rendezvous,... more
In cognitive radio networks, rendezvous is a fundamental operation by which cognitive users establish communication links. Most of existing works consider rendezvous of a pair of users. When multiple pairs of users are doing rendezvous, collisions are caused by the multiple user-pairs which significantly degrade the rendezvous performance, e.g., resulting in a long time-to-rendezvous. To address this problem, we propose a new protocol called Cooperative Rendezvous Protocol which exploits cooperation in the multiple user-pairs environment to speed up the rendezvous operation. Using this protocol, multiple user-pairs cooperate with each other to relay their channel availability information, so that they could avoid attempting rendezvous in unavailable channels. The proposed protocol serves as a general framework that can be applied in conjunction with any existing rendezvous algorithm for faster rendezvous. We theoretically derive an upper bound on the time-to-rendezvous when the proposed protocol is applied in conjunction with any rendezvous algorithm which generates channel hopping sequence based on the available channel set. In addition, we conduct extensive simulation and the results show that the proposed protocol can significantly reduce the time-to-rendezvous of the existing rendezvous algorithms by up to 80.86% in multiple user-pairs environment.
Research Interests:
Research Interests:
Research Interests:
Research Interests:
Research Interests:
Research Interests:
Workshop Chairs Xiaowen Chu, Hong Kong Baptist University, Hong Kong, China Yangdong Deng, Tsinghua University, China Wei Ge, Chinese Academy of Science, China ... Program Committee Tor Aamodt, University of British Columbia, Canada David... more
Workshop Chairs Xiaowen Chu, Hong Kong Baptist University, Hong Kong, China Yangdong Deng, Tsinghua University, China Wei Ge, Chinese Academy of Science, China ... Program Committee Tor Aamodt, University of British Columbia, Canada David Bader, Georgia Institute of Technology, USA Lorena Barba, Boston University, USA John Cavazos, University of Delaware, USA Yifeng Chen, Peking University, China Jack Dongarra, University of Tennessee, USA Geoffrey Fox, Indiana University, USA Markus Hadwiger, King Abdullah ...
Research Interests:
Research Interests:
Indoor robotics localization, navigation and interaction heavily rely on scene understanding and reconstruction. Compared to monocular vision which usually does not explicitly introduce any geometrical constraint, stereo vision based... more
Indoor robotics localization, navigation and interaction heavily rely on scene understanding and reconstruction. Compared to monocular vision which usually does not explicitly introduce any geometrical constraint, stereo vision based schemes are more promising and robust to produce accurate geometrical information, such as surface normal and depth/disparity. Besides, deep learning models trained with large-scale datasets have shown their superior performance in many stereo vision tasks. However, existing stereo datasets rarely contain the high-quality surface normal and disparity ground truth, which hardly satisfy the demand of training a prospective deep model for indoor scenes. To this end, we introduce a large-scale synthetic indoor robotics stereo (IRS) dataset with over 100K stereo RGB images and high-quality surface normal and disparity maps. Leveraging the advanced rendering techniques of our customized rendering engine, the dataset is considerably close to the real-world cap...
Research Interests:
Hyperledger Fabric is a popular open-source project for deploying permissioned blockchains. Many performance characteristics of the latest Hyperledger Fabric (e.g., performance characteristics of each phase, the impacts of ordering... more
Hyperledger Fabric is a popular open-source project for deploying permissioned blockchains. Many performance characteristics of the latest Hyperledger Fabric (e.g., performance characteristics of each phase, the impacts of ordering services, bottleneck and scalability) are still not well understood due to the performance complexity of distributed systems. We conducted a thorough performance evaluation on the first long term support release of Hyperledger Fabric. We studied the performance characteristics of each phase, including execute, order, and the validate phase, according to Hyperledger Fabric’s new execute-order-validate architecture. We also studied the ordering services, including Solo, Kafka, and Raft. Our experimental results showed some findings as follows. 1) The execution phase exhibited a good scalability under the OR endorsement policy but not with the AND endorsement policy. 2) We were not able to find a significant performance difference between the three ordering ...
Research Interests:
The BLAST software package for sequence alignment is one of the most fundamental and widely used bioinformatics tools [1] [2]. Given the large population of BLAST users, any improvement in the execution speed of BLAST will bring... more
The BLAST software package for sequence alignment is one of the most fundamental and widely used bioinformatics tools [1] [2]. Given the large population of BLAST users, any improvement in the execution speed of BLAST will bring significant benefits to the bioinformatics community. Some research groups have used GPUs to accelerate the speed of BLAST. E.g., GPU-BLAST uses GPUs to accelerate BLASTP, and it achieves 3 to 4 times of speedup over single-thread CPU based NCBI-BLASTP [3]. GPUs have also been successfully used to accelerate other sequence alignment tools, e.g., [4]. In this poster, we show our design, implementation, optimization, and experimental results of GPU-BLASTN, a GPU-accelerated version of the widely used NCBI-BLASTN. To the best of our knowledge, this is the first work that provides a complete solution for accelerating BLASTN by GPUs. GPU-BLASTN can obtain identical results as NCBI-BLASTN, and its speed on a contemporary Nvidia GTX680 GPU card is about 10 to 20 ti...
Research Interests:
Multiplication of a sparse matrix to a dense matrix (SpDM) is widely used in many areas like scientific computing and machine learning. However, existing work under-looks the performance optimization of SpDM on modern manycore... more
Multiplication of a sparse matrix to a dense matrix (SpDM) is widely used in many areas like scientific computing and machine learning. However, existing work under-looks the performance optimization of SpDM on modern manycore architectures like GPUs. The storage data structures help sparse matrices store in a memory-saving format, but they bring difficulties in optimizing the performance of SpDM on modern GPUs due to irregular data access of the sparse structure, which results in lower resource utilization and poorer performance. In this paper, we refer to the roofline performance model of GPUs to design an efficient SpDM algorithm called GCOOSpDM, in which we exploit coalescent global memory access, fast shared memory reuse, and more operations per byte of global memory traffic. Experiments are evaluated on three Nvidia GPUs (i.e., GTX 980, GTX Titan X Pascal, and Tesla P100) using a large number of matrices including a public dataset and randomly generated matrices. Experimental ...
Research Interests:
Abstract Conventionally, it is a prerequisite to acquire a good number of annotated data to train an accurate classifier. However, the acquisition of such dataset is usually infeasible due to the high annotation cost. Therefore,... more
Abstract Conventionally, it is a prerequisite to acquire a good number of annotated data to train an accurate classifier. However, the acquisition of such dataset is usually infeasible due to the high annotation cost. Therefore, semi-supervised learning has emerged and attracts increasing research efforts in recent years. Essentially, semi-supervised learning is sensitive to the manner how the unlabeled data is sampled. However, the model performance might be seriously deteriorated if biased unlabeled data is sampled at the early stage. In this paper, an unbiased semi-supervised cluster tree is proposed which is learnt using only very few labeled data. Specifically, a K-means algorithm is adopted to build each level of this hierarchical tree in a decent top-down manner. The number of clusters is determined by the number of classes contained in the labeled data. The confidence error of the cluster tree is theoretically analyzed which is then used to prune the tree. Empirical studies on several datasets have demonstrated that the proposed semi-supervised cluster tree is superior to the state-of-the-art semi-supervised learning algorithms with respect to classification accuracy.
Research Interests:
Rendezvous is a fundamental operation for cognitive users to establish communication links so as to realize data communications and network management. Most of existing rendezvous algorithms implicitly assume that each cognitive user is... more
Rendezvous is a fundamental operation for cognitive users to establish communication links so as to realize data communications and network management. Most of existing rendezvous algorithms implicitly assume that each cognitive user is equipped with one radio, i.e., one wireless transceiver. As the cost of wireless transceivers is dropping, it becomes economically feasible to utilize multiple radios to significantly improve the rendezvous performance. In this paper, we propose an Adjustable Multi-Radio Rendezvous (AMRR) algorithm which exploits multiple radios for fast rendezvous based on available channels only. Suppose that a cognitive user is equipped with m radios. Our basic idea is to partition the radios into two groups: k stay radios and (m-k) hopping radios. The user stays on specific channels in the stay radios while hops on its available channels parallelly in the hopping radios. We prove that the maximum time-to-rendezvous (MTTR) of AMRR is upper-bounded by O(|C_1||C_2|/m_1m_2}), where |C_1| and |C_2| are the numbers of available channels of two users and m_1 and m_2 are the numbers of radios of the two users. This bound meets the lower bound of MTTR of any deterministic rendezvous algorithm when two users are equipped with the same number of radios (i.e., m_1=m_2). AMRR is adjustable in giving its best performance on either MTTR or E(TTR) by adjusting value of k. Simulation results show that AMRR performs better than the state-of-the-art.
Research Interests:
Research Interests:
Research Interests:
... From GOPi+1 to GOPi+3 there are three GOPs that contain the redundant key messages. ... Among these functions, cryptographic hardware accelerators (SHA-1, MD5, DES, 3DES, AES) in NPEB are used in our application for selected video... more
... From GOPi+1 to GOPi+3 there are three GOPs that contain the redundant key messages. ... Among these functions, cryptographic hardware accelerators (SHA-1, MD5, DES, 3DES, AES) in NPEB are used in our application for selected video encryption. ...
Research Interests:
Blocking has been the key performance index in the design of an all-optical network. Existing research demonstrates that an effective routing and wavelength assignment (RWA) strategy and a proper wavelength converter placement algorithm... more
Blocking has been the key performance index in the design of an all-optical network. Existing research demonstrates that an effective routing and wavelength assignment (RWA) strategy and a proper wavelength converter placement algorithm are the two primary vehicles for improving the blocking performance. However, these two issues have largely been investigated separately in that the existing RWA algorithms have seldom
Research Interests:
Rerouting is a viable and cost-effective approach to decrease the blocking probability in legacy circuit-switched networks. We study lightpath rerouting in optical WDM networks in this paper. We investigate two different lightpath... more
Rerouting is a viable and cost-effective approach to decrease the blocking probability in legacy circuit-switched networks. We study lightpath rerouting in optical WDM networks in this paper. We investigate two different lightpath rerouting strategies, namely, passive rerouting and intentional rerouting. Passive rerouting means rerouting established lightpaths to accommodate new lightpath requests which will otherwise be blocked. Intentional rerouting is to intentionally reroute existing lightpaths during their life period without affecting other lightpaths, so as to achieve a better load balancing. Through extensive simulation studies, we draw the following conclusions: 1) when there is wavelength conversion, passive rerouting works much better than intentional rerouting; 2) when there is no wavelength conversion, a naive-wavelength- retuning algorithm can achieve the most benefit of passive rerouting while path-adjusting does not help too much.
Research Interests:
Overlay multicast, which performs topology construction and data relaying in the application layer, has recently emerged as a promising vehicle for data distribution. In most of the existing systems, only a single stream is assumed for... more
Overlay multicast, which performs topology construction and data relaying in the application layer, has recently emerged as a promising vehicle for data distribution. In most of the existing systems, only a single stream is assumed for each overlay, and multiple streams, if needed, are distributed separately. In addition, while the overlay node performs relay functions, they generally do not filter
Research Interests: Computer Networks, Resource Allocation, Broadcasting, Global Optimization, Resource Management, and 13 moreTranscoding, Overlay Networks, Data Distribution, Overlay Network, Bandwidth, Application Layer Multicast, Overlay Multicast, Application Software, Data Stream, Optimal Solution, Generic model, UTILITY FUNCTION, and Hierarchical Structure
Researching aspects of parallel architecture and system design concerning network processors requires excellent simulation tools. In this paper, we develop a tool called NPNS on top of a widely used network simulator ns-2, which combines... more
Researching aspects of parallel architecture and system design concerning network processors requires excellent simulation tools. In this paper, we develop a tool called NPNS on top of a widely used network simulator ns-2, which combines full-system simulators to provide a unified framework for processor and software design within a network context. In this article, we describe the architecture of NPNS and its implementation issues. In addition, an example of simulating a multithreaded Intel IXP1200 network processor with NPNS is illustrated.
Research Interests:
ABSTRACT Recently, sensor data storage has gained increasing popularity for reliable access to data through redundancy spread over unreliable nodes in wireless sensor networks. In storage-centric sensor networks, several schemes have been... more
ABSTRACT Recently, sensor data storage has gained increasing popularity for reliable access to data through redundancy spread over unreliable nodes in wireless sensor networks. In storage-centric sensor networks, several schemes have been proposed to optimize the performance of data management in terms of data availability, repair bandwidth, etc. However, few works have been undertaken to study the performance of these data management schemes from a comprehensive point of view. In this paper, we adopt a concise graphic model, i.e., Stochastic Petri Nets (SPNs), to analyze the performance of three representative data management schemes. From the steady state probability matrix of the SPNs models, we can easily get the average energy consumption, repair bandwidth, reliability and data availability. Based on numerical results, we provide guidelines for designing sensor data storage systems. The results also demonstrate that our proposed models are suitable for analyzing data management schemes in sensor data storage.
Research Interests:
Research Interests: Distributed Computing, Localization, Group communication, Cryptography, Anonymity, and 15 moreEncryption, Authentication, Fault Tolerance, Communication Network, Computer Communications, Reliability, Ad hoc network, Robustness, Data Encryption, Denial of Service, Fault Tolerant, Electrical And Electronic Engineering, Data Transmission, Authentication Protocol, and Denial of Service Attack
This paper presents a fuzzy control approach for process assignment in order to guarantee absolute percentile delay in web servers. Comparing with mean delay used in previous works, percentile delay introduces strong nonlinearities to the... more
This paper presents a fuzzy control approach for process assignment in order to guarantee absolute percentile delay in web servers. Comparing with mean delay used in previous works, percentile delay introduces strong nonlinearities to the plant; it is therefore difficult to achieve satisfactory performance by a linear controller, which is commonly used to guarantee mean delay performance. In this paper we propose a fuzzy controller. Its key feature is independence of the plant and robust to model uncertainty, which is very suitable to the strong nonlinear web servers. The experiments have demonstrated the efficiency of the proposed fuzzy controller in handing the variety of requests.
Research Interests:
Research Interests:
N. Bartolini et al. (Eds.): QShine/AAA-IDEA 2009, LNICST 22, pp. 3451, 2009. © Institute for Computer Science, Social-Informatics and Telecommunications Engineering 2009 ... Risk-Aware QoP/QoS Optimization for Multimedia ... Yanping... more
N. Bartolini et al. (Eds.): QShine/AAA-IDEA 2009, LNICST 22, pp. 3451, 2009. © Institute for Computer Science, Social-Informatics and Telecommunications Engineering 2009 ... Risk-Aware QoP/QoS Optimization for Multimedia ... Yanping Xiao1, Chuang Lin1, Yixin Jiang1, ...
Research Interests:
Research Interests:
Research Interests:
AbstractHomomorphic hash functions (HHFs) have been applied into peer-to-peer networks with erasure coding or network coding to defend against pollution attacks. Unfortunately HHFs are computationally expensive for contemporary CPUs.... more
AbstractHomomorphic hash functions (HHFs) have been applied into peer-to-peer networks with erasure coding or network coding to defend against pollution attacks. Unfortunately HHFs are computationally expensive for contemporary CPUs. This paper proposes to exploit ...
Research Interests:
Research Interests:
Research Interests:
In a heterogeneous peer-to-peer network, different peers provide different qualities of service. It will be very helpful if a peer can identify which peers can provide better services than others. In this paper, we design a novel... more
In a heterogeneous peer-to-peer network, different peers provide different qualities of service. It will be very helpful if a peer can identify which peers can provide better services than others. In this paper, we design a novel reputation model which enables any peer to calculate a reputation value for any other peer that reflects the quality of service provided by