Effective Mapping Of Artificial Neural Network Algorithms Onto Massively Parallel Hardware

Author : Linqian Liu
Publisher :
ISBN 13 :
Total Pages : 168 pages
Book Rating : 4.:/5 (816 download)

DOWNLOAD NOW!

Download or read book A Parallel and Distributed Computing Platform for Neural Networks Using Wireless Sensor Networks written by Linqian Liu and published by . This book was released on 2012 with total page 168 pages. Available in PDF, EPUB and Kindle. Book excerpt: Artificial neural network algorithms inherently possess fine-grain parallelism and offer the potential for fully distributed and local computation. A scalable hardware computing platform that can take advantage of such a massive parallelism and distributed computation attributes of artificial neural networks is considered to be well-poised to compute real-time solution of complex and large-scale problems. This thesis proposes a novel computing architecture for parallel and distributed computation where the hardware-software platform is the wireless sensor networks complete with its wireless protocol stack. More specifically, the proposed idea leverages the existing wireless sensor networks technology to serve as a hardware-software platform to implement and realize certain type of algorithms with fine-grain parallelism, such as those in the domain of artificial neural networks, in massively parallel and fully distributed mode. The research vision is to enable real time computation of solutions of large-scale and complex problems through the proposed parallel and distributed hardware realization of computational algorithms. The thesis defines the new parallel and distributed processing (PDP) and computing architecture and its application for artificial neural network computations. The underlying architectural principles, and structure of the proposed parallel and distributed computing platform are formulated and established. The proposed design is illustrated for feasibility through a simulation-based case study that leverages Kohonen's self-organizing map or SOM neural network on a number of different problem domains or data sets. The research study demonstrates mapping Kohonen's self-organizing map or SOM, configured for a set of domain specific problems, to the proposed PDP architecture. A comprehensive simulation study is conducted to assess the performance profile of and demonstrate the proposed computing architecture, with respect to feasibility. A wireless sensor network simulator (PROWLER) is employed for validation and performance assessment of the proposed computational framework. Three data sets, namely Alphanumeric or Text, Iris, and Wine, where each one differs in the number of attributes, instances, and clusters, are employed to profile the performance of the proposed computing platform. The simulation results are compared with those from the literature and through the MATLAB SOM toolbox. Comparative performance analysis suggests that the proposed computing platform is feasible and promising. The proposed design has potentially much wider applicability for problems with inherent fine-grain parallelism in various domains where mathematics-based problem-solving methodology is not applicable due to lack of a closed-form model for the process or system. Solving complex and very large-scale problems in real time is likely to have radical and ground-breaking impact on the entire spectrum of scientific, technological, economic and industrial endeavors enabling many solutions that were simply not feasible.