Machine Learning 2.0
Imagine a world where the accuracy of machine learning in safety- and security-critical systems could be ensured.
The capabilities of machine learning (ML) continue to improve in everything from detecting cancer to combating malware in computer systems. But ML’s opaque “thinking,” drawn from algorithms used on training data instead of code, presents a particular challenge when a system misbehaves because of a situation outside of its scope—as with recent fatal crashes due to errors in autonomous driving software—or when adversarial actors trick the model.
Junfeng Yang, associate professor of computer science, has spent his career on safety and security for software systems. Now he is building a new tool set for ML.
“Software systems have evolved dramatically and now rely on ML to make many critical decisions,” says Yang. “The traditional tools are no longer adequate to the challenge.”
With fellow computer science professor Suman Jana, he created DeepXplore to test networks in autonomous applications like driverless cars. By computing confusing inputs and feeding them into the network, DeepXplore reverse engineers the learning process to pinpoint where errors occur.
Yang also develops verification tools to prove the absence of bugs and ensure safety. His VeriVis for computer vision systems generates 12 standard image transformations to find misclassified images and verify given safety parameters. In tests with current computer vision systems, VeriVis found thousands of misclassified images and verified the parameters of thousands more.
Yang also looks at data protection and “machine unlearning” to address information in a system that needs to be removed— even from training models—because it’s wrong, it’s poisoned by hackers, or the user desires it for privacy. Machine unlearning removes this data, creating a new model in real time.
Recently, Yang has taken his research in an entrepreneurial direction with NimbleDroid, a platform for businesses that continuously scans mobile apps and immediately fixes slowdowns, crashes, and other critical issues. The service has been praised by clients like Pinterest and the New York Times for improving engineering productivity, user experience, and revenue.