Collaborative research: SHF: small: fault localization for deep learning

The National Science Foundation Division of Information and Intelligent Systems awarded Oakland University $223,992 on June 1, 2026, under the Computer and Information Science and Engineering program (CFDA 47.070) to develop fault localization techniques for deep neural networks.

The project, performed in Rochester, Michigan, through May 31, 2029, will research methods to identify and debug errors in deep learning models during neural network training. Traditional software fault localization techniques do not translate directly to deep neural networks because the underlying computational models and definitions of software defects differ fundamentally. The research will pursue three novel directions: identifying dynamic behavior of DNN models that should be captured in execution traces; defining novel abstractions of dynamic behaviors to enhance fault localization and repair; and developing scalable and generalizable fault localization techniques applicable across different DNN architectures including fully connected neural networks and convolutional neural networks. By enabling earlier error detection during the expensive training phase, the work has potential to reduce training costs and improve the safety and accessibility of AI-based software for practitioners without deep expertise in neural network architecture.

Abstract

This project will study deep learning, a class of machine learning algorithms based on deep neural networks (dnns) that are becoming increasingly popular due to their successful applications in many areas, such as healthcare, transportation, and entertainment. DNN programs, like any other software, may contain faults that might undermine their safety and reliability in mission-critical applications. Software engineering research has produced a rich body of software fault localization techniques; however, they are not immediately applicable to dnns. This is mainly because traditional software and DNN models are based on fundamentally different computational models, and the definition of ?BUG? Differs in the two kinds of software. The project will improve fault localization for dnns with novel approaches for monitoring model behavior during the training of the neural networks. DNN models are also used by practitioners who may not be experts in DNN architecture, and fault localization techniques proposed by this project have the potential to make debugging DNN more accessible, improving the safety and quality of AI-based software. Training DNN models is known to be expensive. This project has the potential to reduce training costs by identifying errors early on that can be rectified. This project will explore three novel research directions. (1) identify dynamic behavior of DNN models that need to be reified in traces. The preliminary work of the investigators has shown that reifying the dynamic behavior of fully connected neural networks (fcnn), such as changes in learnable parameters help with BUG localization in fcnn; however, other model architectures like convolutional neural networks (cnns) have different kinds of learnable parameters. (2) define novel abstractions of dynamic behaviors in DNN models that will enhance fault localization and repair. This research direction will explore the development of new abstractions that can represent the dynamic behavior of the neural networks succinctly. (3) reduce the cost of RE-training for fault localization by leveraging the abstraction of dynamic behavior. In this direction, the project will create abstractions that not only reduce the training dataset but also enhance the effectiveness of fault localization, thereby saving time and computational resources. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.