Recent work in unsupervised feature learning and deep learning has shown that being able to train large models can dramatically improve performance. In this paper, we consider the problem of training a deep network wi...
详细信息
ISBN:
(纸本)9781509007684
Recent work in unsupervised feature learning and deep learning has shown that being able to train large models can dramatically improve performance. In this paper, we consider the problem of training a deep network with hundreds of parameters using distributed CPU cores. We have developed Bagging-Down SGD algorithm to solve the distributing problems. Bagging-Down SGD introduces the parameter server adding on the several model replicas, and separates the updating and the training computing to accelerate the whole system. We have successfully used our system to train a distributed deep network, and achieve state-of-the-art performance on MINIST, a visual handwriting font library. We show that these techniques dramatically accelerate the training of this kind of distributed deep network.
A basic reason for backend systems in mobile application architectures is the centralized management of state. Mobile clients synchronize local states withthe backend in order to maintain an up-to-date view of the ap...
详细信息
We propose an advanced architecture which works better than typical architecture of social networks. We are going to study architecture of social networking sites. Our architecture uses a load balancing module for buf...
详细信息
ISBN:
(纸本)9781450342674
We propose an advanced architecture which works better than typical architecture of social networks. We are going to study architecture of social networking sites. Our architecture uses a load balancing module for buffering user queries before transmitting them to data server. In this paper, query optimization of user queries for faster results has been discussed. Our experimentation exhibits possibility of data failure reduction due to external disturbances.
the integration network of terrestrial and satellite is becoming a new trend in the development of network applications in the future. Software defined network and virtualization is becoming a hot research topic and a...
详细信息
ISBN:
(纸本)9781509051540
the integration network of terrestrial and satellite is becoming a new trend in the development of network applications in the future. Software defined network and virtualization is becoming a hot research topic and application of the ground network. It is necessary to study the application of software defined network and virtualization in satellite applications. this paper discusses the software defined networking and network virtualization in satellite network application system architecture. Software defined satellite network will separate the control plane and data plane of satellite network and can on the whole manage and configure the satellite network resources from the whole network so as to realize the optimization of satellite network. the system architecture is described and the design scheme of virtual satellite network is shown. In this paper, the design principle of the satellite OpenFlow switch and controller, the realization methods of network function virtualization and the controller are described in detail. Based on this kind of system architecture, the global dynamic load balancing, routing and resource scheduling algorithms are designed, and the performance analysis and simulation of network delay and throughput are carried out.
Spiking neural P systems(in short, SNP systems) are parallel and distributedcomputing devices inspired by the function and structure of spiking neurons. Recently, a new variant of SNP systems, called SNP systems with...
详细信息
ISBN:
(纸本)9781509039067
Spiking neural P systems(in short, SNP systems) are parallel and distributedcomputing devices inspired by the function and structure of spiking neurons. Recently, a new variant of SNP systems, called SNP systems with structural plasticity(in short, SNPSP systems) was introduced. In SNPSP systems, neuron can use plasticity rules to create and delete synapses. In this work, we consider many restrictions sequentially on SNPSP systems:(i) we use the weighted synapses;(ii) neuron withthe maximum number of spikes is chosen to fire. Specifically, We investigate the computational power of weighted SNPSP systems working in maximum spiking strategy(in short, WSNPSPM systems) and we proved that such SNPSP systems are universal as generating devices and accepting devices.
the results of the statistical model experiments research of various load balancing algorithms in distributedcomputing systems are presented. Software tools were developed. these tools, which allow to create a virtua...
详细信息
the results of the statistical model experiments research of various load balancing algorithms in distributedcomputing systems are presented. Software tools were developed. these tools, which allow to create a virtual infrastructure of distributedcomputing system in accordance withthe intended objective of the research focused on multi-agent and multi-threaded data processing were developed. A diagram of the control processing of requests from the terminal devices, providing an effective dynamic horizontal scaling of computing power at peak loads, is proposed.
the proceedings contain 195 papers. the topics discussed include: detecting pedestrians behavior in building based on Wi-Fi signals;e-health services for elderly care based on Google cloud messaging;on interesting pla...
ISBN:
(纸本)9781509018932
the proceedings contain 195 papers. the topics discussed include: detecting pedestrians behavior in building based on Wi-Fi signals;e-health services for elderly care based on Google cloud messaging;on interesting place finding in social sensing: an emerging smart city application paradigm;collaborative representation-based robust face recognition by discriminative low-rank representation;a self-adaptive video dual watermarking based on the motion characteristic and geometric invariant for ubiquitous multimedia;the role of big data on smart grid transition;sensing information modeling for smart city;exploration of collective pattern to improve location prediction of mobile phone users;performance modeling of preemption-based packet scheduling for data plane in software defined networks;and aspect-oriented customization of the scheduling algorithms and the resource access protocols of a real-time operating system family.
Recommender systems are used to filter through vast amounts of items and recommend those that potentially have the highest relevance for the user. Recently, research dealing with recommendations in academia increased....
详细信息
Modern distributed multimedia applications are resource-hungry, and they often leverage on-demand cloud services to reduce their expenses. Existing cloud services deploy many servers in a few data centers, which consu...
详细信息
ISBN:
(纸本)9781509014453
Modern distributed multimedia applications are resource-hungry, and they often leverage on-demand cloud services to reduce their expenses. Existing cloud services deploy many servers in a few data centers, which consume a lot of electricity to power up and cold down, and thus are expensive and environmentally unfriendly. In this paper, we present a multimedia fog computing platform that utilizes resources from public crowds, edge networks, and data centers to serve distributed multimedia applications at lower costs. We use animation rendering as a case study, and identify several challenges for optimizing it on our multimedia fog computing platform. Among these challenges, we focus on the problem of predicting the completion time of each rendering job. We propose an efficient algorithm based on stateof-the-art machine learning algorithms. We also fine-tune the algorithm using multi-fold cross-validation for higher prediction accuracy. With real datasets, we conduct trace-driven simulations to quantify the performance of our prediction algorithm and that of the whole platform. the simulation results show that our proposed algorithm outperforms a state-of-the-art statistical model in several aspects: completed job ratio by 20%, makespan by 2 times, and normalized deviation by 30 times, on average. Moreover, the overall performance of the platform with our proposed algorithm is fairly close to that with an Oracle of the actual job completion time: a small factor of 1.48 in terms of makespan is observed.
Motivated by an attempt to understand the formation and development of (human) language, we introduce a "distributed compression" problem. In our problem a sequence of pairs of players from a set of K player...
详细信息
ISBN:
(纸本)9783959770293
Motivated by an attempt to understand the formation and development of (human) language, we introduce a "distributed compression" problem. In our problem a sequence of pairs of players from a set of K players are chosen and tasked to communicate messages drawn from an unknown distribution Q. Arguably languages are created and evolve to compress frequently occurring messages, and we focus on this aspect. the only knowledge that players have about the distribution Q is from previously drawn samples, but these samples differ from player to player. the only common knowledge between the players is restricted to a common prior distribution P and some constant number of bits of information (such as a learning algorithm). Letting Tϵ denote the number of iterations it would take for a typical player to obtain an ϵ-Approximation to Q in total variation distance, we ask whether Tϵ iterations suffice to compress the messages down roughly to their entropy and give a partial positive answer. We show that a natural uniform algorithm can compress the communication down to an average cost per message of O(H(Q) + log(D(P||Q)) in ∼O (Tϵ) iterations while allowing for O(ϵ)-error, where D(||) denotes the KL-divergence between distributions. For large divergences this compares favorably withthe static algorithm that ignores all samples and compresses down to H(Q) +D(P||Q) bits, while not requiring Tϵ K iterations that it would take players to develop optimal but separate compressions for each pair of players. Along the way we introduce a "datastructural" view of the task of communicating with a natural language and show that our natural algorithm can also be implemented by an efficient data structure, whose storage is comparable to the storage requirements of Q and whose query complexity is comparable to the lengths of the message to be compressed. Our results give a plausible mathematical analogy to the mechanisms by which human languages get created and evolve, and in particular highl
暂无评论