Federal Project Grant Award Summary

Tulane University received a $376,008 Project Grant from the National Science Foundation's Division of Information and Intelligent Systems under the Computer and Information Science and Engineering program (CFDA 47.070) for collaborative research on fault localization in deep learning systems. The grant, awarded June 1, 2026, with a completion date of May 31, 2029, will develop novel approaches and techniques to identify and locate faults in deep neural networks (DNNs). The research addresses a critical gap in software debugging capabilities, as traditional fault localization methods applicable to conventional software cannot be directly applied to DNN models due to their fundamentally different computational architectures and the distinct definition of "bugs" in machine learning contexts.

The project will deliver three primary research contributions: (1) identification of dynamic DNN behaviors requiring detailed monitoring and tracing during neural network training, with particular focus on extending preliminary findings from fully connected neural networks to other architectures such as convolutional neural networks; (2) development of novel abstractions of dynamic behaviors that will enhance both fault localization and repair capabilities for DNN models; and (3) additional research directions to be explored during the three-year performance period. Deliverables from this research have the potential to reduce DNN training costs by enabling early error detection and correction, improve accessibility of DNN debugging for non-expert practitioners, and enhance the safety and reliability of artificial intelligence-based software applications across mission-critical domains including healthcare, transportation, and entertainment.