Making AI Sleep More Like Humans
Electrical Engineering and Computer Science Professor Garrett Rose is part of an interdisciplinary team of researchers that was awarded a $2 million grant from the National Science Foundation (NSF) Emerging Frontiers in Research Initiatives program.
The grant will fund research to bridge the gap between human brain processing efficiency and the limitations of current artificial intelligence (AI) models. This endeavor seeks to create a new form of AI that rapidly learns, adapts, and operates in uncertain conditions, all while effectively addressing the energy challenge plaguing modern AI.
This project helps the team of researchers embark on rapidly advancing frontiers of fundamental engineering research and address a grand challenge in AI. University of Texas, San Antonio’s Dhireesha Kudithipudi, an expert in brain-inspired AI and Energy-efficient AI, will spearhead this ambitious research project as the lead Principal Investigator.
Rose is part of an accomplished team of co-principal investigators who will contribute their expertise to various facets of the research. Rose brings neuromorphic computing expertise to the team. The other team members are:
- Assistant Professor of Psychology Itamar Lerner (UTSA) specializes in the theory of Neuroscience.
- Associate Professor of Computer Science Christopher Kanan (University of Rochester) brings deep learning insights.
- Associate Professor of Philosophy John Basl (Northeastern University) serves as the senior personnel responsible for AI ethics.
The team will draw ideas from “Temporal Scaffolding Hypothesis,” a theory that mirrors the human brain’s ability to process temporal patterns during both wakefulness and sleep. Unlike contemporary AI models, the human brain achieves this with remarkable energy efficiency, processing diverse temporal information across varying timescales. This stark contrast serves as the driving force behind the quest to create AI models that can emulate the human brain’s adaptability and efficiency.
According to this hypothesis, the brain can capture temporal patterns effectively by reactivating wake experiences during sleep in an accelerated manner, enabling the detection of crucial temporal patterns within these experiences. The team will develop deep learning networks and bio-plausible spiking neural networks that mimic this behavior and assess their efficiency on large temporal problems. Once developed, such networks will have the ability to swiftly adapt and operate under resource constraints, much like the human brain. This transformative approach has the potential to tackle AI challenges in learning and energy consumption, with applications envisioned in critical sectors such as healthcare, autonomous systems, and national security.
UT will be involved in all the circuit design aspects for the project. Rose and his staff will be using their existing hardware and inputting spiking models to test them.
“Something I am hoping is that we can get at some of the questions related to what biology is doing with spiking information during these sleep phases,” Rose said. “If that is useful for people and even mice to retain information, maybe it’s useful for artificial systems.”
In addition to advancing AI technology, this project seeks to empower the future AI workforce. The team is committed to providing new training opportunities for underrepresented students in AI fields, fostering a competitive AI workforce to maintain U.S. technological leadership in science, technology, engineering, and mathematics (STEM).
“I am very excited about working with this team to unlock the potential of AI,” Rose said. “Having Tennessee play a role in this project demonstrates the great opportunities available for our staff and students to make positive contributions to pioneering research in the world of engineering.”