Parallel backpropagation neural networks forTask allocation by means of PVM

• Autores: Crespo, María Liz; Printista, Alicia Marcela; Piccoli, María Fabiana
• Año de publicación: 1998
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias Informáticas; Informática; Distributed Systems; System architectures; Neural nets; PATTERN RECOGNITION; system architecture; distributed systems workload; parallelised neural networks; backpropagation; partitioning schemes; pattern partitioning
• Institución: Universidad Nacional de La Plata

Features such as fast response, storage efficiency, fault tolerance and graceful degradation in face of scarce or spurious inputs make neural networks appropriate tools for Intelligent Computer Systems. A neural network is, by itself, an inherently parallel system where many, extremely simple, processing units work simultaneously in the same problem building up a computational device which possess adaptation (learning) and generalisation (recognition) abilities. Implementation of neural networks roughly involve at least three stages; design, training and testing. The second, being CPU intensive, is the one requiring most of the processing resources and depending on size and structure complexity the learning process can be extremely long. Thus, great effort has been done to develop parallel implementations intended for a reduction of learning time. Pattern partitioning is an approach to parallelise neural networks where the whole net is replicated in different processors and the weight changes owing to diverse training patterns are parallelised. This approach is the most suitable for a distributed architecture such as the one considered here. Incoming task allocation, as a previous step, is a fundamental service aiming for improving distributed system performance facilitating further dynamic load balancing. A Neural Network Device inserted into the kernel of a distributed system as an intelligent tool, allows to achieve automatic allocation of execution requests under some predefined performance criteria based on resource availability and incoming process requirements. This paper being, a twofold proposal, shows firstly, some design and implementation insights to build a system where decision support for load distribution is based on a neural network device and secondly a distributed implementation to provide parallel learning of neural networks using a pattern partitioning approach. In the latter case, some performance results of the parallelised approach for learning of backpropagation neural networks, are shown. This include a comparison of recall and generalisation abilities and speed-up when using a socket interface or PVM.
Sistemas Inteligentes
Red de Universidades con Carreras en Informática (RedUNCI)