Machine learning is starting to creep into all sorts of applications in our everyday lives, from search engines and digital assistants to traffic alerts. In the near future, even more sensitive parts of our lives will incorporate machine learning, including self-driving cars, financial institutions and medical devices and procedures.
But while current machine learning models are great at some tasks, they are not foolproof. That’s why Grigorios Chrysos, who joined the University of Wisconsin-Madison Department of Electrical and Computer Engineering as an assistant professor in fall 2023, is working on reliable machine learning, or ways to improve the privacy, safety and robustness of artificial intelligence.
“As we get ready to launch some of these technologies more widely, we need to take great care and study their properties closely in the years to come,” he says. “For instance, if you try to launch a radiology program developed in the United States in a country that does not speak English, or where the format of the machines is not the same, can you guarantee it will be safe and sound? What will its output be? These are the types of things we need to investigate.”
Grigorios began programming computers as an early teen and studied electrical engineering and computer science at the National Technical University of Athens in Greece, earning his bachelor’s and master’s degrees.
As a PhD student at Imperial College London, Grigorios started out studying computer vision—in particular, how to classify deformable objects, which are objects that can change shape, like human hands. During that research, however, he began thinking about neural networks, a method of artificial intelligence in which computers process data in a way similar to the human brain.
“This was around 2018, and people were very excited, saying deep neural networks will solve everything,” he says. “I was like, okay, great, they might be able to solve everything. But why? Why is this a unique way to solve everything? Or can we find other types of functions that can solve everything? Neuroscientists are still debating if neural networks reflect what the brain is really doing. So why did we put all our eggs in a single basket?”
Grigorios began investigating other potential AI models. “In the end, with a bit of luck and a bit of curiosity, we ended up with something called polynomial networks,” he says.
This type of artificial intelligence, based on the mathematical concept of polynomial expansions, is well-suited to some applications, including a popular face-generation tool. As a postdoctoral scholar at EPFL in Lausanne, Switzerland, over the last three years, Chrysos has continued work on polynomial networks, extending the theory behind the model.
At Madison, he hopes to continue that work, as well as to focus on reliable machine learning. That means developing theory and tools to make sure artificial intelligence doesn’t go off the rails. Even in systems that work 99.9% of the time, he says, extreme edge cases can cause problems. He points to recent incidents in which chatbots released to the public began spewing hate speech or told people in mental distress to harm themselves.
“We need to make sure all the models we launch are fail-safe,” Grigorios says. “If we want to launch them to make our life easy, we need to ensure that they make life easy for everybody.”
Madison is a great place to continue his work, and not just because it already has a strong core of machine learning researchers in engineering and the computer sciences. Grigorios says he was attracted by the university’s commitment to recruiting talent from around the world. As a longtime member and former chairman of the Electrical Engineering Students’ European Association, he has had the opportunity to connect with engineers in many European countries. That type of cross-cultural interaction, he says, has been a huge part of his personal and academic development. UW-Madison’s diverse faculty, he says, has a similar feel. “I appreciate places with a tradition of fostering academic ideas, even beyond my field,” he says. “Because this means they are the perfect environment to nurture new ideas and create the future.”
Top photo by Joel Hallberg