An infant’s brain is able to gradually learn and adapt, often more easily than an adult brain. For example, research shows that before adolescence, children are able to pick up languages much faster than adults. Portland State’s Christof Teuscher uses this same analogy to describe new artificial intelligence (AI) technology he helped develop.
“The problem with current artificial neural networks is that they’re very static, built for a very specific problem, and trained with a very specific dataset,” said Teuscher, a PSU professor in Electrical and Computer Engineering. “What we developed as a breakthrough in AI technology is a novel type of device and approach to build systems that can be completely changed and reconfigured on the fly for different applications. We can, for example, create neurons on synapses on demand, as the system needs to learn new things.”
The breakthrough in technology was developed in partnership with Purdue University, Pennsylvania State University, Santa Clara University, Argonne National Laboratory, University of Illinois, Brookhaven National Laboratory, University of Georgia and University of Illinois at Chicago. An article discussing their findings, “Reconfigurable perovskite nickelate electronics for artificial intelligence,” was published in the February issue of Science.
Current artificial intelligence systems suffer from what’s known as catastrophic forgetting. When an existing system is programmed to learn something new, it forgets what it had already learned, forcing the system to essentially start from scratch.
To better conceptualize the applications of this research, Teuscher offered Apple’s Siri as an analogy.
“Siri can speak English, French, German, you name it, right? It has an existing source of knowledge,” he said. “Now we can use our platform to teach a completely new language without Siri forgetting all the other languages that it already speaks.”
Further, this new system can be reconfigured for all types of applications using existing neural network building blocks and create a system that gradually grows and evolves, as the demand and its environment change.
“This opens up avenues for AI technology that continuously learns, grows as needed, and gradually improves, something that current AI systems simply can’t do. It’s a very important step towards more general, more open-ended, more capable AI,” Teuscher said. “It’s a technology that has huge potential to improve many aspects of our lives.”