Aditya Thakur, an assistant professor in the Department of Computer Science, has received an NSF CAREER Award for his proposal, “Provable Patching of Deep Neural Networks.”
The National Science Foundation’s Faculty Early Career Development Award (CAREER), is the agency’s most prestigious award for early-career faculty. It recognizes young researchers who have the potential to serve as role models and leaders in research and education at their institutions or in their fields. The award supports research projects that can serve as the foundation for the rest of their careers. Thakur is one of six College of Engineering faculty to receive a CAREER Award this year.
"I am honored to receive an NSF CAREER award,” said Thakur. “This grant will enable us to continue our work on foundational problems in safety and fairness of artificial-intelligence systems.”
Thakur's project tackles the problem of repairing mistakes in deep neural networks (DNNs). DNNs are a type of machine learning algorithm loosely modeled after the human brain, which aim to mimic human intelligence by learning from experience. DNNs are comprised of a vast, layered web of “neurons” that are trained using data to make complex, probability-based decisions.
DNNs have been successfully applied to a wide variety of problems, including image recognition, natural-language processing, medical diagnosis and self-driving cars, but as the accuracy of these networks has increased, so has their complexity and size. Moreover, DNNs are far from infallible, and mistakes they make can have disastrous consequences. Thakur’s goal is to reduce or eliminate this danger by developing techniques and tools to repair a trained DNN once a mistake has been discovered.
The project introduces the notion of provable patching, which involves making a minimal change to the parameters of a trained DNN to correct its behavior according to a given specification. To do this, he will leverage his experience in the areas of formal methods and machine learning to develop theoretical foundations, design efficient algorithms and evaluate practical applications of provable patching for DNNs.
The project builds on Thakur’s previous work with DNNs, formal methods and program verification, which has been funded by Facebook and been featured at top conferences in software engineering and machine learning. With the CAREER Award, he plans to build on this success and make an impact on both the field and the students he teaches and mentors.