A Posit based Handwritten Digits Recognition System

Abstract

This paper presents a novel design of a Posit-based handwritten digits recognition system, one of the convolutional neural network applications. Herein, LeNet and ResNet-18 based HDRS (Handwritten Digits Recognition System) architecture is used for training and inference of model. The parameters obtained after training were converted to (8, 0) Posit number system. Training of LeNet and ResNet-18 based HDRS has been done over the MNIST database, an open-source database for handwritten digits recognition. The proposed Posit (8, 0) based HDRS provides comparable accuracy to traditional single-precision floating point and fixed point based HDRS.