Training neural network as approximate 4:2 compressor applying machine learning algorithms for accuracy comparison

Abstract

This research work is on interlinking the emerging application namely machine learning algorithm and an approximation technique to compare the accuracy and loss values of the trained machine learning based approximate 4:2 compressor model in accordance with the proposed approximate 4:2 compressor such that the trained model is sufficient to observe the performance of the approximate model without performing error analysis separately using MATLAB. The key attention of this work is to adopt truth tables from an exact and proposed 4:2 compressor as data set and train the artificial neural network as the respective 4:2 compressor model. The Neural Networks are trained for two data sets, firstly by applying inputs and only Sum output and second with inputs and three (Cout, Carry and Sum) outputs. Error Analysis of approximate 4:2 compressor has been performed with MATLAB and training of artificial neural network has been done employing Anaconda Python with Jupyter Integrated Development Environment. A comparison of error rate has been made for exact and approximate 4:2 compressors and accuracy has been evaluated for trained neural networks (as exact and approximate 4:2 compressors) and found that the loss values are smaller and the difference between the trained accuracy values are less for trained and validated proposed approximate machine learning model.