Mizzou Engineers Take Transfer Learning Step Further in New Paper
Imagine having a self-driving car that is already trained to recognize road markings, street signs and other vehicles. Could you take what that car knows and tweak it so the vehicle could navigate the ocean, too? The idea is known as transfer learning, and a Mizzou Engineering team has taken it a step further by applying the technique to the organization, or clustering, of data.
“One of the big ideas of the new wave of artificial intelligence is this idea of transfer learning,” said Jim Keller, Curators’ Distinguished Professor in the Department of Electrical Engineering and Computer Science. “You’ve got one group of data where you know a lot about it — either you have millions of images or some underlying knowledge about the problem itself. What you want to do is investigate another set of data where you don’t have a lot of knowledge about it.”
In a recently published paper, Keller and his team applied the concept to an algorithm that clusters information by type. For instance, the self-driving car may know how to organize objects as street signs but would not know how to cluster underwater obstacles into categories such as fish or coral reefs.
Applying transfer learning to clustering has numerous potential real-world applications. For the paper, researchers looked at data sets they had from sensors used in eldercare settings. An eldercare facility could have millions of points of data from existing patients, for instance, that could be used to help organize preliminary sensor data from a new patient.
“In eldercare, you might have some residents you’ve seen for five years and you’ve got a lot of data for them,” Keller said. “So how do you use that to help you with somebody who just showed up where you don’t have a lot of data? Can you use that data to help you better understand new data quicker? What this paper does is demonstrates that you can use the results of the clustering in the source (or original) space to help guide you to the correct clustering in the target (or new) space.”
The paper, “Transfer Learning Possibilistic C-Means,” was published in the In Institute of Electrical and Electronics Engineers (IEEE) Transactions on Fuzzy Systems and promoted in the IEEE’s Computational Intelligence Society’s May newsletter.
Co-authors were Rayan Gargees, a PhD student in electrical and computer engineering, and Mihail Popescu, a professor of health management and informatics with a joint appointment in electrical engineering and computer science.