Evaluation of Posit Arithmetic on Machine Learning based on Approximate Exponential Functions

Recent advances in semiconductor technology lead to ongoing applications to adopt complex techniques based on neural networks. In line with this trend, the concept of optimizing real number arithmetic has been raised. In this paper, we evaluate the performance of the noble number system named posit on neural networks by analyzing the execution of approximate exponential functions, which is fundamental to several activation functions, with posit32 and float32. To implement the functions with posit arithmetic, we designed the software posit library consisting of basic arithmetic operations and conversion operations from/to C standard data types. The result shows that posit arithmetic reduces the average relative error rate by up to 87.12% on the exponential function.