ASE 2023 Distinguished Paper Award for ISU Research on Fault Localization of Deep Neural Network

September 20, 2023

Dr. Ali Ghanbari, currently a postdoctoral research associate in the Department of Computer Science at Iowa State University, and Dr. Hridesh Rajan, Kingland Professor and Chair of the Department of Computer Science, have received the Distinguished Paper Award at the 38th IEEE/ACM International Conference on Automated Software Engineering (ASE 2023) for their paper titled "Mutation-based Fault Localization of Deep Neural Networks." ASE conference is a top-tier venue for research on techniques for automating software engineering practices such as software bug detection and localization.

This work applies mutation-based fault localization to deep neural networks. “This research is an example of SE4AI, i.e., Software Engineering for Artificial Intelligence. It revisits mutation-based fault localization in the context of deep neural network models and proposes a novel technique, named deepmufl, applicable to a wide range of deep neural network models,” explained Ghanbari.

Mutation-based fault localization is based on mutation analysis, which involves systematically introducing small, controlled variations or mutations into a program's code to observe how these alterations impact the program's behavior. By employing this methodology within the domain of deep neural networks, Ghanbari has crafted a versatile bug detection technique, named “deepmufl.” This technique contributes to state-of-the-art by detecting more bugs from a sizable bug dataset.

More Reliable Deep Neural Networks

Deep neural networks (DNNs) represent a remarkable advancement in artificial intelligence. They function as highly sophisticated programs capable of emulating human-like cognitive abilities. DNNs are particularly adept at recognizing images, understanding complex human language, and making intricate decisions.

Researchers have recently introduced tools and techniques to help identify and correct bugs in DNNs. In this research, Ghanbari introduces "deepmufl," a mutation-based fault localization technique specialized for DNNs, helping pinpoint where bugs might be within the network. He has applied deepmufl on 109 buggy DNN programs obtained from StackOverflow, a popular coding website. Impressively, deepmufl successfully detected 53 out of these 109 bugs. This outperforms existing DNN fault localization techniques regarding the number of detected bugs, making it a promising solution for enhancing the reliability of DNN-based systems.

Ghanbari added, “This work lays the foundations for future research on automated debugging of data-driven systems.”

As DNNs and other complex software systems continue to grow, and complexity, the need for faster and more effective bug-detection techniques becomes increasingly apparent. Therefore, building upon the success of deepmufl, there are plenty of opportunities for future research on improving the speed of mutation-based fault localization for DNNs, and fixing the located bugs automatically.

The Critical Importance of Addressing Bugs in Deep Neural Networks

Like any software, though, DNNs are not immune to bugs, which can have far-reaching consequences when considering the critical roles that DNNs play in various applications. These neural networks can encounter various problems, significantly impeding their performance and functionality.

In the context of DNNs, these bugs can manifest in various ways. For example, in the case of image recognition, a DNN might misclassify objects in images, leading to errors in tasks like identifying pedestrians, vehicles, or obstacles. In natural language understanding, bugs could cause the DNN to misinterpret language, potentially resulting in miscommunications or incorrect responses. Moreover, in decision-making scenarios, DNNs might make erroneous choices based on input data, which can be detrimental in critical applications such as autonomous vehicles or medical devices.

Given the expanding role of DNNs in domains like autonomous transportation and healthcare, addressing these bugs becomes vital. Imagine the potential consequences of a self-driving car misinterpreting a traffic signal or failing to detect an obstacle on the road. Such errors could lead to accidents and put lives at risk. Similarly, in the medical field, if a diagnosis tool powered by a DNN provides incorrect information or misses critical details, it could impact patient outcomes and safety.

Finding and rectifying these bugs is paramount in scenarios where human lives and safety are at stake. Ensuring the reliability and trustworthiness of DNN-based systems is not just a matter of convenience but life and death. Consumers and society must be able to rely on these systems to perform flawlessly, especially when applied in situations where human error or system. Failure could result in catastrophic consequences. Therefore, pursuing robust and bug-free DNNs is not merely an option but an ethical and practical imperative in our increasingly interconnected and AI-driven world.

News Type: