A research project at Indiana University Bloomington will receive support from the Center for Quantum Technologies, the country's only National Science Foundation-funded center that partners academic researchers with private sector tech companies to advance quantum technology.
As a part of the Center for Quantum Technologies, IU researchers in computational linguistics and mathematics at the College of Arts and Sciences at IU Bloomington will explore how to significantly reduce the time and energy costs required to train artificial intelligence in understanding language.
"Artificial intelligence is currently hugely expensive in terms of time, money and energy," said project co-leader Damir Cavar, an associate professor in the College's Department of Linguistics. "We're seeing reports right now about tech companies lobbying the government to extend the lifetime of nuclear power plants simply to guarantee them enough power to train their AI models.
"Quantum computing has the potential to be far more energy efficient because you're basically computing on the atomic level. That's what we're working on."
The co-leader on the project is Larry Moss, a professor and associate chair in the College's Department of Mathematics and a top expert in the mathematics of logic and reasoning. Also funded under the project is Rong Zheng, a Ph.D. student in the Department of Psychology.
As one of several new projects named under Year 2 of the Center for Quantum Technologies, IU researchers will meet regularly with representatives from the public and private sector to share research results and receive guidance on their work. These partners include the Air Force Research Laboratory; Amazon Web Services; Cummins; D-Wave Systems; Eli Lilly and Co.; Entanglement Inc.; Peraton; and Quantum Corridor.
According to Cavar, the "semantic mapping" required to generate AI models is the central challenge faced by every company using the technology. Describing these equations as "advanced linear algebra," Cavar said they are the foundation of the technology that powers AI's ability to accept input that resembles normal human speech and output results in a similar manner - a concept known as "natural language processing."
As the founder of IU's computational linguistics program, Cavar has spent more than 20 years studying the equations required to convert words - as well as images and concepts - into "vectorized representations." For example, he said, the vector that represents the word "apple" and the vector that represents the word "tree" have similar distances and spatial relations in semantic space to the vectors that represent the word "grape" and "vine," reflecting analogy of meaning in multi-dimensional space.
But the machine learning processes required to convert billions upon billions of words, images and concepts into mathematical abstractions is massively expensive in terms of time and energy, he said.
Under the new project, Cavar, Moss and a small cadre of graduate students, including Zheng, are working to convert existing AI algorithms for natural language processing from traditional software coding environments to quantum computing environments. The team is testing is equations on real quantum hardware using cloud-based services such as the IBM Quantum Platform and Amazon's Braket.
Most of the team's brainstorming work occurs each week in the basement of IU Bloomington's Luddy Hall through meetings of the Quantum Natural Language Processing Study Group, Cavar said. The group is already readying three academic papers for publication based on the results of the sessions.
However, Cavar added that converting existing algorithms from traditional computing environments to quantum environments is merely a "proof of concept" compared to advancements on the horizon. He said the "real next leap" in the use of quantum computers for artificial intelligence is leveraging their unique properties to produce results that have not yet been achieved with traditional computers, such as on-the-fly reasoning and greater information reliability.
"Even if you're converting vector and matrix-based operations to a quantum environment, it's still linear algebra," he said. "But in the quantum world, you can represent concepts in a more complex way - with a potentially infinite number of states that they can represent. It gives you a much broader instrument with which to represent semantics and meaning."
The result could potentially produce a system more able to engage in logic and reasoning, he said. A system that doesn't produce "hallucinations" - including answers that sound accurate but contain misinformation - but rather a system that understands concepts deeply enough to eliminate logical errors before delivering a response.
"That's the next frontier: more advanced systems that are able to reason based on new information or plan out a logical series of actions - and that don't require a whole nuclear power plant's worth of energy to do it," Cavar said.
The Quantum Natural Language Processing Study Group
