In this paper we try to identify the most promising way to execute the training and the testing of a Multi Layer Perceptron neural network, by considering the use of a Matlab implementation on a PC and an embedded architecture based either on a general purpose multiprocessor or on a dedicated device. We compared the performance obtained from these three different approaches by first implementing a classification problem of recognising the region of the space to which randomly extracted points belong to and then facing a biomedical signal fitting problem. In both applications the superiority of the dedicated chip in terms of performance was evident. Moreover the dedicated chip uses a training technique, known as Reactive Tabu Search algorithm, which is often more efficient than the Back Propagation approach used in the other solutions.
A parallel neural processor for real-time applications
DANESE, GIOVANNI;LEPORATI, FRANCESCO;RAMAT, STEFANO
2002-01-01
Abstract
In this paper we try to identify the most promising way to execute the training and the testing of a Multi Layer Perceptron neural network, by considering the use of a Matlab implementation on a PC and an embedded architecture based either on a general purpose multiprocessor or on a dedicated device. We compared the performance obtained from these three different approaches by first implementing a classification problem of recognising the region of the space to which randomly extracted points belong to and then facing a biomedical signal fitting problem. In both applications the superiority of the dedicated chip in terms of performance was evident. Moreover the dedicated chip uses a training technique, known as Reactive Tabu Search algorithm, which is often more efficient than the Back Propagation approach used in the other solutions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
