Six UT faculty and students received inaugural funding from the UT Austin-Amazon Science Hub for their research on artificial intelligence and machine learning.
Created in April, the Science Hub is a five-year collaborative effort between the University and Amazon to support research in topics ranging from machine learning to networking and communications.
Greg Durrett, an associate computer science professor, was awarded $75,000 to continue his work in natural language processing — the systems that allow computers to understand human language.
“These days, with the development of things like ChatGPT, we’re looking a lot at the capabilities of large language models … particularly focusing on their ability to conduct complex reasoning tasks, and then how we can think about making their outputs truthful,” Durrett said.
Durrett’s research uses these large language models to verify the accuracy of other language models.
“Large language models are some of the best tools that we have for (fact-checking) because it’s not a simple process of looking it up in a database,” Durrett said. “That’s kind of the broad goal here, to build the system that can go all the way from some text produced by a model to make sure that everything it says is factual.”
Durrett said the funding will help support further research into the accuracy of AI language models’ outputs.
“It’s been much talked about how these systems don’t always generate the right stuff,” Durrett said. “They might just quote unquote, hallucinate facts, or generally kind of stitch things together in ways that may misrepresent the sources, and so this kind of stuff is only useful insofar as we can trust it.”
Georgios Smyrnis, an electrical and computer engineering doctoral student, received funding for his work on helping computers distinguish between unlabeled data.
“Say that you have an image of a cat and an image of a dog,” Smyrnis said. “In these paradigms, you give the model the images of the cats and dogs, but you never explicitly tell them which is which, so this way, you may use techniques that allow you to differentiate between the data without actually knowing what the data means or where the data is.”
Smyrnis said his research has a range of applications outside of machine learning. The award will help him create smaller models and fund the computer needed for this research.
“At the end of the day, what is important about creating smaller models for this project is to make them easy to use by pretty much everyone,” Smyrnis said. “The way it stands now, a major bottleneck to using this type of model is how costly they are to us, so by making them smaller and easier to use, we hope to make them more accessible.”
Durrett said the UT Austin-Amazon Science Hub will advance research at UT by leveraging Amazon’s leading innovations in language and dialog processing, which the company uses to develop products like Alexa.
“There’s a lot of mutual benefit that we can have by further collaboration between UT and Amazon,” Durrett said.
