A0819
Title: Experimental design and active learning
Authors: Rong Pan - Arizona State University (United States) [presenting]
Abstract: In machine learning or artificial intelligence, supervised learning methods such as classification and regression are so important that almost 80\% ML/AI practice is about supervised learning. To perform supervised learning, one must have labeled data to build and train the learning model. However, labeling data are often expensive, while unlabeled data are cheap to obtain. Also, in specific tasks, not all available data are equally useful. The question is how to find the good, useful data to label them at minimal cost, while receiving maximum benefit, so as to learn the system more efficiently. Parallel to this notion, statistical experimental design is about deriving a strategy of selecting experimental conditions to conduct experiments such that the expected experimental results can best achieve the experimenter's objective. Therefore, both AL andDOE concern how to take samples from a population. We will present some recent developments of active learning (AL) in the ML/AI field and draw the connection of AL to traditional experimental design methodologies, particularly optimal design and sequential design. We will discuss how optimal design and sequential design theories can provide some theoretical enhancements to AL as well as practical improvements of AL algorithms.
