Research team wins award for best innovative paper
By Bridget Maiellaro, ECE Illinois
November 20, 2007
- Prof. Thomas Huang and grad students Yun Fu and Xi Zhou recieved the 2007 DoCoMo USA Labs Innovative Paper Award from IEEE.
- The paper, entitled “LAPLACIAN Affinity Propagation for Semi-Supervised Object Classification,” focuses on machine learning and pattern recognition.
- The group accomplishes this by solving semi-supervised multi-class object classification problems with the use of a graph-based learning algorithm known as the Laplacian Affinity Propagation (LAP).
ECE students Yun Fu and Xi Zhou, in collaboration with ECE Professor Thomas S. Huang, recently received the 2007 DoCoMo USA Labs Innovative Paper Award at the IEEE International conference on Image Processing.
The paper, entitled “LAPLACIAN Affinity Propagation for Semi-Supervised Object Classification,” focuses on machine learning and pattern recognition. The group accomplishes this by solving semi-supervised multi-class object classification problems with the use of a graph-based learning algorithm known as the Laplacian Affinity Propagation (LAP).
“We developed a general framework which could reformulate most graph-embedding algorithms for semi-supervised learning,” Fu said. “We further developed a particular algorithm based on this framework and used it to deal with object classification and also clustering applications.”
Fu said the concept of the paper stemmed from semi-supervised learning, which is the idea of using unlabeled training samples to help design classifiers. The paper included research results on 20 classes of object images, including cars and toys. Some of the images were for training, while others were for testing.
“Such label-to-unlabel propagation scheme can provide a closed form solution via a graph-based learning framework that is flexible for any new design,” according to the abstract of the paper. “LAP integrates embedding and classifier together and gives smooth labels with respect to the underlying manifold structure formed by the training data.”
Huang said that labeling objects by human is normally very tedious, and, therefore, the algorithm is beneficial because it saves time.
“Let’s say you have many images of dogs and then you have many images of cats, and you label them dog and cat,” Huang said. “So the system will look at the label training sample to learn what makes a dog a dog and what makes a cat a cat. Therefore, when you have an unlabeled image coming in, then the algorithm will decide whether it is a dog or a cat.”
Of the 1,708 papers submitted, 843 were accepted. DoCoMo USA Labs, a subsidiary of NTT DoCoMo USA, Inc., awarded only two of those papers. The conference was held from Sept. 16 to Sept. 19 at the Hyatt Regency San Antonio in Texas. Fu gave a 20-minute presentation on the group’s paper on Sept. 17. Fu said that he talked about the paper’s potential applications and explained the basic idea, while the audience asked questions. Zhu Li, a student at Northwestern and employee of Motorola Labs who Fu met during his summer internship in 2006, also contributed to the research and paper.
As for now, the researchers are working on a variety of things. Fu has worked with Professor Huang for three years and plans to graduate next year. Eventually he hopes to become a professor at a university. Zhou, whose research focuses on human-computer interaction and machine learning, still has a few years before he graduates.
Editor's note: media inquiries should be directed to Brad Petersen, Director of Communications, at email@example.com or (217) 244-6376.