Machine Learning using Logarithmic Arithmetic with Preconditioned Input to Mitchell's Method

Mitchell's method often has approximated base-two antilogarithms using low-cost hardware in low-accuracy systems. Another application of identical circuits is to approximate the the addition-logarithm function needed for the sum of values represented by the Logarithmic Number System (LNS). This paper shows how preconditioning the input to Mitchell's method improves the accuracy of Mitchell logarithmic addition and subtraction at lower cost than other methods. One promising application is machine learning. The proposed preconditioning has MNIST training performance nearly identical to full FP.