Computer Science Professor Baishakhi Ray Wins NSF Career Award

Ray seeks to improve the reliability of deep-learning applications used in safety-critical systems such as autonomous cars, medical diagnosis and malware detection

Apr 23 2019 | By Bernadette Young | Photo Courtesy of Baishakhi Ray

Baishaki Ray

The National Science Foundation (NSF) has awarded Baishakhi Ray, assistant professor of computer science, the NSF CAREER Award, its most prestigious honor given to early-career faculty. The five-year, $530,143 award will support Ray’s research proposal, “Systematic Software Testing for Deep Learning Applications.” Her research seeks to improve the reliability of deep-learning applications, with possible implications on safety, security, and fairness of these applications.

Ray aims to develop a technique to systematically test deep-learning-based applications—those used in safety-critical systems including autonomous cars, medical diagnosis, malware detection, and aircraft collision avoidance systems. The technique will also help core machine learning experts to evaluate their models in a systematic fashion.  

Ray’s work addresses a paradigm shift in software development, where decision making is increasingly shifting from hand-coded program logic to deep learning as its core component. Such a wide adoption of deep-learning techniques comes with concern about the safety and reliability of these systems, which have experienced a number of reported erroneous behaviors.

For her project “Systematic Software Testing for Deep Learning Applications,” Ray plans to devise novel software testing strategies for deep-learning applications and to build and evaluate a framework, DeepTest, to realize those strategies. In particular, the proposed framework will develop novel white-box testing strategies, realistic test-case generation techniques, and regression-testing techniques. If successful, DeepTest will help software developers to evaluate the reliability of the underlying applications and detect buggy behaviors during the development and maintenance phase.

A unique characteristic of deep neural network applications is that they depend highly on training data. The logic of the model can change drastically if it is trained using a different data distribution. Consequently, the reliability of the model greatly depends on how representative the training data is to the real world. “I think it is important that while testing a deep neural network application, we should test both the model and the training data,” said Ray, who joined Columbia Engineering last fall.

“With this NSF support I want to make machine learning systems safer and robust,” said Ray. “I believe systematically testing deep-learning applications is an emerging and important software engineering problem, especially given their increasing deployment in safety-critical systems.”