Researchers suggest that a recent hypothesis on how the human brain uses alternating sleep and wake periods for continuous learning might hold the solution to addressing the challenges of lifelong learning in artificial intelligence.
Christopher Kanan, an associate professor at the University of Rochester‘s Department of Computer Science, is amongst a cross-disciplinary team awarded $2 million in funding from the National Science Foundation (NSF). The team aims to apply the ‘temporal scaffolding’ hypothesis to develop AI that quickly learns, adapts, and functions effectively in unpredictable conditions.
Existing AI models face challenges in continuously learning new tasks over a lifetime and often exhibit subpar performance in environments with limited resources. Kanan highlights that although models have advanced in sequence learning where lessons from past experiences guide future decisions, they still struggle to match the efficiency of human learning.
He observes that traditional reinforcement learning approaches resulted in AI models capable of outperforming professional players in games like StarCraft 2 and Dota 2. However, these AI models required extensive experience to attain such proficiency.
“OpenAI’s Dota 2 AI required the equivalent of 45,000 human years’ worth of gameplay experience to beat the world’s best players, but the best players only had been playing the game a few years at the time,” says Kanan. “This indicates something is fundamentally wrong with how reinforcement learning works. Instead, we are proposing to use an alternative paradigm to hopefully get much more efficient learning.”
The hypothesis of temporal scaffolding suggests that the brain rapidly reactivates wake experiences during sleep, allowing it to discern crucial patterns within those experiences. In emulating this process, the team aims to create deep-learning networks capable of rapid adaptation and functioning under resource constraints, mirroring the efficiency of the human brain.
Kanan will create deep-learning models and algorithms grounded in the temporal scaffolding concept, subjecting them to standardised benchmarking tasks. The principal investigator, Dhireesha Kudithipudi from the University of Texas at San Antonio (UTSA), and co-PI Garret Rose from the University of Tennessee, Knoxville, will also endeavour to implement selected algorithms developed by Kanan’s lab into hardware.
The temporal scaffolding hypothesis, inspiring these algorithms, originated from UTSA neuroscientist Itamar Lerner, while Northeastern University’s John Basl will assume the role of senior personnel overseeing AI ethics. This project represents Kanan’s most recent venture in the evolving realm of artificial intelligence and deep learning.
“The one thing I’m sure of is ChatGPT and others like it are here to stay,” says Kanan. “And we, as educators, will just have to deal with that.” He predicts that these innovations will significantly disrupt higher education and many other settings.