Singh Ready to Prove Deep Neural Networks Can Be Safe and Reliable Through NSF CAREER Award

Five years ago, Illinois Computer Science professor Gagandeep Singh began branching out from this his initial research emphasis to work on the reliability and safety of Deep Neural Networks (DNNs).

Despite enjoying previous projects focused on a problem considered intractable for more than 40 years – the design of precise and scalable numerical program analysis – Singh has identified DNNs as an area of emphasis gaining traction in the artificial intelligence community. This is due to many academicians, he said, realizing the limitations of deep learning following years of excitement and promise.

In fact, it may well be his research background that will serve as a stabilizing force to the future of DNNs, considering he has developed mathematically rigorous solutions that also provide practical systems useful to society.

Certainly, the NSF valued Singh’s experience when considering his recent proposal – entitled “Proof Sharing and Transfer for Boosting Neural Network Verification” – that earned the NSF CAREER Award earlier in February.

“Verifying the safety and robustness of DNNs is one of the main problems in modern machine learning,” Singh said. “DNN verification is a fundamentally hard problem, and scaling to large realistic DNNs is the main challenge. State-of-the-art methods could only verify DNNs with a few hundred neurons five years ago, while some of the latest verifiers can handle DNNs with up to a million neurons.

“Despite these advances, existing verifiers are fundamentally inefficient when used inside industrial DNN development pipelines that require running the verifier hundreds of thousands of times for different networks and specifications.”

Singh expounded upon the topic by noting that the inefficiency he’s found is that the “verifier starts from scratch for every new pair of networks and specifications.”

Working alongside students through his FOCAL Lab@UIUC, Singh noticed that this process could benefit from incremental verification powered by proof sharing and transfer.

The resulting promise of this work, has Singh thinking big picture.

“DNNs are currently the dominant AI technology and could potentially have a transformative impact on society and the economy. However, these gains will only be realized if they are perceived as safe and reliable,” Singh said. “For establishing trust, we need formal guarantees on DNN behavior in unseen scenarios. The inefficiency of existing verifiers hinders their use in real-world settings.

“I anticipate that the methods and systems developed as part of this project will accelerate the adoption of formal verification within the DNN development and deployment pipelines in diverse industries, including agriculture, computing, finance, and healthcare.”

Singh’s confidence in the work ahead stems from the acceptance of past work.

Beginning from the time he earned ACM’s SIGPLAN Doctoral Dissertation Award, Singh has found great value in building his methodology upon mathematically rigorous and sound theory. The work he has done in the past also developed efficient and easy-to-use systems that other researchers enjoy building upon.

Taking a similar approach to this NSF CAREER Award funded project should provide further proof that he’s on the right path with his work – a sense both friends within academia, and event outside of it, are starting to understand.

“The NSF CAREER Award is prestigious and highly competitive; I’m appreciative of receiving it, as the award is vindication of the effort we’ve put in,” Singh said. “On a personal level, my friends outside of computer since also know of the NSF CAREER Award, even more so than the SIGPLAN Doctoral Dissertation Award. So, now they are more convinced than ever that I am doing something good!”