A Fractional Powers-Of-Two Number System For Digital Neural Networks
Jerry J. Cupal, Bogdan M. Wilamowski, Richard S. Sandige, Julie Miller · International Journal of Modelling and Simulation · 1996
A new arithmetic number system based on fractional powers of two is proposed. In this system, the weight of each bit position is a power of 21/n, where n = 1,2,3,4 ... Multiplication or division by 21/n can be accomplished by simple shifts of the input data. Although not as simple as in conventional twos complement binary, addition and subtraction can be accomplished using simple ALU structures. Details of how this is accomplished are given, as well as an implementation of hardware structure to perform multiply-accumulate operations commonly found in neural nets and FIR filters. Applications of this arithmetic system in binary neural networks are also given.