MIT Technology Review’s annual list of Innovators Under 35 recognizes individuals who are tackling hard problems and making notable advances in the areas of AI, virtual reality, robotics and security. This year’s list included Ian Goodfellow, the inventor of generative adversarial networks, and Franziska Roesner, who focuses on augmented reality (AR) security. Recipients are given an opportunity to make a short presentation of their work at Tech Review’s annual EmTech conference in Cambridge, Mass. And here’s what Goodfellow and Roesner — whose research was selected here for their enterprise applicability — had to say.
Ian Goodfellow & generative adversarial networks
One pain point for machine learning is the man hours it takes to label training data. “At Street View in 2013, we built the first system that was able to read with super human accuracy,” said Ian Goodfellow, staff research scientist at Google Brain, the search engine’s deep learning research project. “And we did that using a database of over 10 million labeled photos.”
In 2014, Goodfellow set out to make this process more efficient. Instead of relying on traditional model optimization to minimize the difference between what a model predicts and what a human says a model should predict, Goodfellow opted for game theory. He calls the technique generative adversarial networks (GANs), and it’s creating a lot of buzz in AI circles.
Two unsupervised neural networks are pitted against each other in a zero-sum game “to learn everything they can about the distribution generating the training data,” Goodfellow said. One model, the discriminator, tries to figure out if an image is real or artificially created. The other model tries to generate realistic-looking images to trick the discriminator model into thinking they’re real. The game is played until the discriminator model cannot distinguish real from artificially-created images.
The neural networks teach each other without the use of labeled data, and the results so far are promising. GANs that used 100 images of digits achieved the same level of accuracy as traditional models trained on 60,000 images just a few years before.
Franziska Roesner & AR security
The benefits of AR tech are easy to dream up: They can overlay instructions on how to fix a sink, provide digitally generated directions onto physical roads, and act as a guide when cooking dinner.
“But there’s also a potential dark side,” said Franziska Roesner, assistant professor of computer science and engineering at the University of Washington. What happens, she asked, when a user accidentally installs a malicious app that can deliberately distract or even obscure objects from sight? Or when users learn that their AR apps or platforms are recording and analyzing their every move?
“My work addresses the critical gap with emerging augmented reality technologies — protecting security, privacy and safety of end users,” she said.
Roesner and her team started by identifying and classifying potential risks and vulnerabilities — from data sensor protection and security to safety risks from virtual content. She and her team then began working on an AR platform that can mitigate vulnerabilities by, say, enforcing safety policies on the virtual content that applications try to display, she said.
Her research is only the beginning, but paying attention to these vulnerabilities while the technology is still not widely deployed is critical, Roesner argues. “My work lays the foundation to enable these technologies to reach their full potential, unhampered by security, privacy and safety risks,” she said.