Methodology for Determining Optimal Model and Training Data in Deep Learning

Abstract

Deep Learning has numerous applications in various market segments including consumer, industrial, energy & utilities, oil & gas, surveillance, autonomous vehicles and medical and so on. And within each dataset, the deep learning inference precision achieved with a trained model may be meeting the precision goals. But that does not mean the same trained model may perform equally well on some other test dataset. In practical applications, such drop in precision on test data variations is highly undesired. Hence, it is critical to train the deep learning model with adequate and augmented training data. Also, it is important to deploy an optimal deep learning model for a given application. This is to utilize optimum compute resources, when such deep learning trained model is deployed in inferencing. This becomes even more important for resource constrained and battery-operated embedded edge applications. Hence, determining the amount of training data needed and deep learning model to be used should not be on trial-and-error basis. There are no known structured methodologies available, for optimal model selection and training data. In this paper, a methodology has been proposed for determining optimal deep learning model and training data to be used, for achieving target precision levels.