Inside A Deep Learning Library, the Symptoms, Causes, and Fixes for Bugs

Abstract

Deep learning has become a popular study area in recent years. While deep learning software generates tremendous positive results in some cases, faults in the programme can have fatal repercussions, in safety-critical applications, this is extremely significant. Researchers have conducted various empirical investigations on defects in deep learning systems to better understand the bug characteristic of deep learning software. Despite the fact that these studies are useful, none of them investigate the problem in TensorFlow, a deep learning framework. They feel that some fundamental concerns about deep learning library difficulties remain unsolved. The answers to these questions are crucial and helpful because they constitute the underlying library of many deep learning endeavours. Many deep learning initiatives are influenced by its flaws. In this paper we pre-process a dataset that is openly accessible by performing data normalization and log transformation. To prepare the data input for the deep learning model, we next undertake data modelling. Third, we send the modelled data to a deep neural network-based model that was specifically created to forecast the number of faults. We test the suggested method on two well-known datasets as well. The evaluation's findings show that the suggested strategy is reliable and can outperform cuttingedge methods. The suggested approach enhances the squared correlation coefficient by more than 8% while, on average, drastically reducing the mean square error by more than 14%.