Professor of Computer Engineering and Computer Science and Director of the Multimedia Research Lab
at the University of Louisville, KY
- Title of the talk:
Multiple Instance Learning with Applications
In standard supervised learning, each object is represented by a single feature vector and a label. This label is categorical for classification problems and real-valued for regression problems.
However, for some applications, providing a label to each observation is either impossible or too tedious and does not scale to big data. These applications cannot be solved effectively using the traditional learning paradigm. An alternative framework of learning that tackles the inherent labeling ambiguity better than supervised learning is the multiple instance learning (MIL) paradigm. In MIL, an object is represented by a collection of feature vectors, or instances, called a bag. Each bag can contain a different number of instances. Labels are available at the bag level, however, labels of individual instances within a bag are unknown. This many-to-one relationship between instances and data labels produces an inherent ambiguity in determining which instances in a given bag are responsible for its associated label.
In this talk, I will describe many multiple instance classification and regression algorithms. Various applications will be used to illustrate the assumptions made by these algorithms and their performance. These applications include: detection of buried explosive objects using 3D ground penetrating radar image data; classification of lung nodules in thoracic screening 3D CT images; multi-sensor information fusion; and image annotation.
- Short biography:
Hichem Frigui is a Professor of Computer Engineering and Computer Science and Director of the Multimedia Research Lab at the University of Louisville, KY. He has been active in the research fields of machine learning; data mining; audio, image, and video processing; and multi- sensor information fusion. He has developed and applied algorithms from the above fields to a wide range of applications including: detection of buried explosive objects using ground penetrating radar and wideband electro-magnetic induction sensors; segmentation and analysis of sports video; Analysis of video simulating medical crises; Analysis of exhaled breath data and computed tomography (CT) images for the diagnosis of lung cancer; Analysis of stable isotope assisted metabolomics (SIAM) data to facilitate the identification of metabolites in biochemical pathways; and Computational tools for predicting material properties; Dr. Frigui’s research has been funded by the National Science Foundation (NSF), Office of Naval Research (ONR), Army Research Office (ARO), and Kentucky Science and Engineering Foundation (KSEF). He has coauthored over 210 technical publications. In 2002, he has received the NSF CAREER award for outstanding young scientists. He served as an Associate Editor of Fuzzy Sets and Systems journal and the IEEE Transactions on Fuzzy Systems.