Adopting Misclassification Detection and Outlier Modification to Fault Correction in Deep Learning-Based Systems

Abstract

Over the past few decades, researchers in software engineering (SE) have focused on testing, analyzing, repairing, and generating programs automatically and effectively. Today, combining neural networks and traditional software engineering techniques has major potential to benefit software quality and productivity. Regarding the development of neural networks, deep learning (DL) and convolution neural networks (CNNs) have been widely adopted by software applications for making decisions or providing suggestions. Considering life-critical DL-based applications, there is a need to correct the wrong decisions made by DL systems immediately. Therefore, we propose a novel fault-correction framework for alleviating potential misclassification issues of DL systems called the Outlier Modification for DL Systems (OMDLS). Our experiment results with two public datasets using different scales and label numbers to show that modifying the outliers based on the misclassification pairs can improve accuracy by up to 2.12% without retraining the model and modifying the inference immediately.