Abstract
Is it possible for multiple machine learners to learn collaboratively, by teaching each other without human intervention? Combining ideas from active learning (learners select items to query for a new label) and co-training (learners label data for each other), we have developed collaborative learning algorithms that enable a group of learners to autonomously improve their individual and joint performance. This approach is ideal for settings such as sensor networks, in which each learner (at each sensor node) obtains a different view on common events of interest. The complementary information enables learners to provide relevant feedback to each other. In supervised collaboration, the learners query and receive feedback in the form of labels for individual data items. In unsupervised collaboration, the learners exchange pairwise constraints on item cluster memberships. In both cases, we observe performance improvements when the learners are able to collaborate. As in co-training, overfitting can happen and therefore early termination may be desirable. We applied collaborative learning to seismic sensor data collected at the Mount Erebus Volcano Observatory in Antarctica, with the goal of separating eruption events from icequakes, and found modest improvements in classification accuracy (from 70% to 80%) and dramatic improvements in clustering accuracy (from 0.2 to 0.6 ARI). This is joint work with Jillian Green (California State University, Los Angeles) and Umaa Rebbapragada (Tufts University).

Bio
Kiri Wagstaff is a senior researcher in artificial intelligence and machine learning at the Jet Propulsion Laboratory. Her focus is on developing new machine learning and data analysis methods, particularly those that can be used for in situ analysis onboard spacecraft (orbiters, landers, etc.). She has developed several classifiers and detectors for data collected by instruments on the EO-1 Earth orbiter, Mars Pathfinder, and Mars Odyssey. The applications range from detecting dust storms on Mars to predicting crop yield on Earth. She has also worked on system modeling and automatic code generation for the Electra radios used by the Mars Reconnaissance Orbiter and Mars Science Laboratory. She holds a Ph.D. in Computer Science from Cornell University and an M.S. in Geological Sciences from the University of Southern California.