Experimental study on the precision requirements of RBF, RPROP and BPTT training
Urs Vollmer · 1999
Most neurocomputer architectures support only fixed point arithmetic which allows a higher degree of VLSI integration but limits the range and precision of all variables. Up to now the effect of this limitation on neural network training algorithms has been studied only for standard models like SOM or BP. This paper presents the results of an experimental study in which the precision requirements of three other learning algorithms (RBF, RPROP and BPTT) on exemplary tasks have been investigated. While the RBF and BPTT key variables required more than 16 bit for training to solve the selected problems, the RPROP algorithm showed good results with far less than 16 bit. 1 Introduction Today neural networks are widely and successfully used to tackle a variety of classification and approximation problems. Since their learning algorithms are computationally expensive and the network architecture suggests a parallel implementation, generalpurpose parallel computers and neurocomputers (like C...