PhD student Divake Kumar and colleagues publish article in Nature Communications

Nature logo

Associate Professor Amit Ranjan Trivedi and several of his students published an article in Nature Communications, one of the most prestigious and highly selective journals in science and engineering. Trivedi’s student Divake Kumar was a first-year PhD student when the article was published. Kumar and Trivedi’s former students Leila Rahimifard and Priyesh Shukla coauthored the paper with their South Korean counterparts.

The paper “Highly parallel and ultra-low-power probabilistic reasoning with programmable gaussian-like memory transistors,” was published this spring.

Deep learning has been integrated into artificial intelligence and is used in areas such as autonomous driving, image recognition, and translation. Environmental factors, including fluctuations in light, limited resolution, and constraints on sensor range, inject uncertainty into the accuracy of the models. The team used probabilistic interference procedures to overcome this uncertainty.

Kumar has a background as a hardware engineer and wants to work on algorithms to optimize hardware for particular applications. He came to UIC in 2023 and joined Trivedi’s lab.

“If you go beyond the current AI models, which are basically working with complete information, you realize there is a need for new kind of higher level intelligence, which can work with incomplete, partial, or even unreliable information,” Kumar said. “That’s where this reasoning capability comes in – these models should be able to self-criticize; they should be able to have this iterative loop of understanding and decision making where they can  actually judge their own predictions and refine these to produce something they’re confident with.”

For wearable devices or cell phones, this work is time consuming and will use a large amount of computing power. So, in collaboration with the South Korean researchers, they developed a type of reasoning transistor, which allows these reasoning-based algorithms to be implemented more efficiently.

 “It’s a little bit more of an artistic way of doing intelligence, where you are not just doing a feed forward, or one-shot prediction, but rather you have a loop going on through which you are analyzing or critiquing things, and covering all the angles before making a prediction,” Kumar said.

Trivedi received a supplemental NSF CAREER award  that will allow his lab to further their research into integrating multiple modes of intelligence to improve efficiency.