Feed-Forward ANN for 2-1 Fixed Point ALUs

S. Vassiliadis K. Bertels G. G. Pechanek E.E. Dept Dept. Gestion de l'Entreprise DSP Architecture/Development T.U. Delft, Delft Univ. of Namur IBM Corporation The Netherlands 2628 CD 5000 Namur, Belgium RTP, N.C. 27709 USA Abstract In this paper we investigate the possibility of constructing xed point units using feed-forward neural networks. We investigate the possibility of constructing small depth neural networks for operations usually de ned in general purpose computer architectures. In particular we show that xed operations require no more depth than the networks for binary addition. We show that depth-3 networks with bounded weights and small size requirements can be constructed that guarantee architectural compliance for xed point arithmetic and address generation operations. Consequently, it is suggested that the proposed scheme can be used to potentially produce high performance arithmetic devices for xed point processing units with small size requirements.